We've been waiting on Nvidia's long-rumored N1X Arm chips for a while at this point. Through several leaks and official teases, the company's ARM-based consumer SoC has excited many as it's poised to open the gates of high-end Arm performance on Windows machines. Interestingly, it seems like the "Chinese Nvidia" has beaten the Green Team to the punch with its own custom Arm chip in a new laptop.

Moore Threads, the region's local darling, has just launched the "MTT AI Book" — a new thin-and-light laptop powered by an in-house "MT1000" CPU. What's special about this chip is that it's Arm-based and features 12 CPU cores clocked at 2.65 GHz (base), along with an unknown GPU that's based on its MUSA microarchitecture. The NPU is capable of delivering up to 50 TOPS of AI compute, similar to AMD's Strix Point.

The SoC as a whole is called "Yangtze" (translated) and is paired with 32 GB of LPDDR5X-7500 unified memory, meaning that's shared between the MT1000 CPU and the MUSA-based GPU. The 1 TB SSD onboard carries a Linux-based operating system called "AIOS," but the device can also run Windows. That's what makes this exciting, because now we're in N1 territory. Unfortunately, we're not looking at Windows-on-Arm here, but rather a virtualization-based approach where Windows just sits inside a VM.

The Nvidia N1 silicon is said to have a 20-core ARM CPU and an RTX 5070-level GPU because Jensen Huang himself confirmed it powers the GB10 Superchip inside the DGX Spark.

Even if we imagine for a second that Moore Thread's Arm SoC is natively running Windows and is optimized — it just doesn't stack up to what Nvidia is cooking. The Green Team's offering is supposed to open up AI and gaming in a whole new way for Windows-on-Arm. Meanwhile, Qualcomm is already trying with its own X series of SoCs.

AMD and Intel, on the other hand, don't have competing Arm products so we can't really speculate much. The Red Team's Strix Halo chips, which feature desktop-level integrated graphics, and the rumored Nova Lake-AX (cancelled?) lineups are both technically in the spot that N1X might gun for: a powerful, portable machine with strong battery life.

Moore Thread's Yangtze doesn't seem to be there yet, and that's proven by its Geekbench listing: it scores 1,127 points in the single-core test and 7,420 points in multi-core. Those numbers are very underwhelming. The most modern CPU we could find around these results was the Ryzen 3 7320U at 1,112 single-threaded points; even a recent Core i3/Core Ultra 3 SKU scores more than that.

Apart from its intriguing silicon, the MTT AI Book features a 2.8K 14-inch OLED display running at 120 Hz. The port selection seems to be limited to just 3x USB-C ports, and the battery is rated at 70Wh. The laptop weighs 1.5 kg despite being CNC-milled out of a "6-series" aluminum alloy. It also looks very similar to a MacBook Air and is priced at 9,999 CNY on JD.com, or about $1,447 USD.

We hope to see more mainstream media coverage of this device with independent reviews that test the Arm-based SoC's capability. This was just one Geekbench listing, so there's still a chance that with the right drivers and firmware tuning, the MTT AI Book can deliver better performance. It's aimed at AI applications, though, so we may not be that impressed by its graphical prowess when it surfaces.

Even though Nvidia's AI GPUs and rack-scale solutions remain the most sought-after AI accelerators, curbs set on exports of Nvidia's AI processors to China, first by the White House and then by Beijing, are having a drastic effect on the company's presence in the People's Republic. As a result, the company's share in China could drop to just 8% in the coming years as domestic suppliers can satisfy around 80% of local demand, reports Nikkei, citing analysis from Bernstein.

"The new products meet the needs of domestic developers," said Zhang Jianzhong, chief executive of Moore Threads, at a news conference while announcing the codenamed Huashan product, the company's first GPU dedicated solely for the acceleration of AI workloads. "There will be no more need to wait for advanced products from overseas."

Analysts from Bernstein cited by Chinese media expect Nvidia's share of China’s AI processor market to drop to around 8% this year from 66% in 2024 as Huawei, Cambricon, and other local independent hardware vendors (IHVs) together approaching 80%. The rise of Chinese hardware accelerators is a result of a combination of events, including restrictions set on Nvidia hardware, progress of hardware from companies like Huawei, Cambricon, Moore Threads, and MetaX, as well as substantial improvements in their software stacks.

Moore Threads' Huashan can compete against Nvidia's Hopper H100 and H200 products, the company's previous-generation AI accelerators that the U.S. recently allowed to export to China, but with some serious strings attached. However, they are considerably slower than Nvidia's existing Blackwell B200 and B300 GPUs, which are barred from export to the People's Republic.

Meanwhile, Huawei's AI CloudMatrix 384 can beat both GB200 NVL72 and GB300 NVL72 systems in BF16 FLOPS, a popular format used for AI training, albeit with four times more power consumption. The company's next-generation Atlas 950 SuperCluster, based on 524,288 Ascend 950DT AI accelerators, is projected to offer up to 524 FP8 ExaFLOPS for AI training and up to 1 FP4 ZettaFLOPS for AI inference (MXFP4 to be more specific) sometimes in 2026 – 2027 and 4 ZettaFLOPS by the end of 2028. This is still behind leading Blackwell-based clusters, such as Oracle's OCI Supercluster running 131,072 B200 GPUs and offering peak performance of up to 2.4 FP4 ZettaFLOPS for inference, but it is evident that Chinese developers are rapidly increasing the performance of their AI hardware.

Given the progress, the remaining hurdle is completing the transition from an ecosystem long centered on Nvidia to a fully domestic hardware and software stack, which may not be that easy to achieve, as many existing AI deployments use Nvidia hardware and Nvidia CUDA software stack and porting them to Chinese hardware and software is hard and expensive.

Yet, transition to domestic AI hardware (and domestic hardware in general) is China's long-term national goal. A draft five-year plan reportedly circulated by the Communist Party in October calls for semiconductor self-reliance under a 'new national system' that directs state bodies, private companies, and financial institutions. At the heart of this effort are the so-called 'four little dragons' of Chinese GPUs: Moore Threads, MetaX, Biren Technology, and Suiyuan Technology (Enflame).

Large hyperscalers are also intensifying their custom silicon programs. Baidu's Kunlunxin unit plans to introduce five AI processors by 2030, and Alibaba is also not giving up on its own silicon efforts. Yet, to a large degree, China's AI industry is limited by SMIC's ability to produce chips on its 7nm-class process technologies in sizable quantities. If the company cannot increase its output substantially in the coming years, then either China's AI sector will fall behind America's dramatically, or it will find a way to obtain high-performance GPUs from Nvidia to keep up.

Chinese GPU maker Moore Threads just held its MUSA Developer Conference, where it unveiled "Huagang," its next-gen architecture set to debut next year. Huagang will span gaming and AI, with significant performance gains promised for both. The event was light on details, so we don't actually have any specs; just claims of what to expect, and we have a lot to look forward to if these promises are true, such as a 15x uplift in "AAA" gaming and a whopping 50x boost in ray tracing performance.

Let's start with "Lushan," the new gaming GPU powered by Huagang that will succeed the existing MTT S80 and S90 models. The latter is the best GPU Moore Threads has offered for a while, and it barely outperforms the RTX 4060, so a rehaul was long overdue. With Lushan, Moore Threads is claiming a 15x uplift in "AAA" gaming — whatever that means (raster?) — and a whopping 50x boost in ray tracing performance.

Moreover, the company is quoting a 64x increase in AI compute, 16x in texture geometry processing, 4x in texture fill, 8x in atomic access, and 4x in memory capacity. For context, the S80 and S90 carry 16 GB of GDDR6 VRAM, so next year's Lushan GPUs are expected to feature up to 64 GB of memory, which would be interesting to see in these times.

Another vital improvement to Moore Threads' new GPUs is full support for modern APIs such as DirectX 12 Ultimate, which should alleviate compatibility concerns. There is a dedicated 2nd-gen hardware ray tracing engine, along with a new AI hardware block for the "UniTE" unified rendering architecture, which should bring the GPU's rendering pipeline to parity with offerings from Nvidia, AMD, and Intel.

Alongside Lushan, Moore Threads also teased the "Huashan" AI GPU, which consists of two chiplets and 8 HBM modules. The performance is being touted as comparable to Nvidia's Hopper and Blackwell GPUs, with memory bandwidth exceeding even the B200. The company also claims a 50% increase in compute density and a 10x improvement in efficiency.

Huashan supports FP4 through FP64 compute and offers exclusive low-precision mixed formats: MTFP4, MTFP6, and MTFP8. In terms of connectivity, Moore Threads plans to scale these GPUs across AI factories, with over 100,000 GPUs interconnected at up to 1314 GB/s via the MTLink 4.0 interconnect.

While we didn't see any gaming or AI benchmarks for these next-gen GPUs, Moore Threads demonstrated DeepSeek V3 performance on the MTT S5000 GPU, achieving 1000 tokens per second in Decode and 4000 tokens per second in Prefill. These results put it slightly ahead of Nvidia's Hopper lineup, which has historically been the cap for Green Team in China in terms of AI GPUs.

The MTT S5000 will launch next year, but it's not part of the Huashan lineup, as it's been in the news before. We should learn more about MT's Huagang GPUs in the coming months as the company takes industry giants like Nvidia, Intel, and AMD head-on with a state-backed Chinese alternative that should help fuel the region's self-reliance ambitions.

A sudden flood of blacklisted Chinese chipmakers in both the GPU and CPU arenas are preparing to IPO over the coming months. The development comes as the rise of AI and U.S. sanctions against China's semiconductor and supercomputer sectors have had a drastic effect on the development of China's semiconductor industry.

On the one hand, companies like Huawei initiated projects to develop advanced fab tools that could eventually replace equipment produced in the U.S. and Europe. On the other hand, over half a dozen GPU companies emerged to develop processors for AI and gaming. Coincidentally, they are all looking forward to going public and making plans for their IPOs.

There are multiple private GPU developers in China, including Biren Technology, DenglinAI, Innosilicon, Illuvatar CoreX (Tianshu Zhixin), MetaX Technology, Moore Threads, and Zhaoxin. Some of these companies — Innosilicon, Moore Threads, and Zhaoxin — are formally focused on gaming GPUs (though it does not mean that their GPUs cannot be used for AI or other data center applications), others — Biren, DenglinAI, Illuvatar CoreX, and MetaX — primarily develop processors for AI and supercomputer applications.

For now, all of these semiconductor companies are private, but for many, this will change soon.

Some Chinese GPU companies pursue IPO

Perhaps the most promising China-based GPU developer, Biren, remains privately held but has been actively working toward becoming a public company for the last couple of years. It is undergoing the required regulatory guidance process in China, a prerequisite before listing. The timeline and venue (such as Shanghai's STAR Market or Hong Kong) have not been finalized publicly.

Moore Threads is also still private, but it has taken formal steps toward a domestic IPO. In late 2024, it restructured into a joint-stock company, a standard step for firms preparing to go public in mainland China. It is expected to list on one of China's major exchanges, pending regulatory approval.

China’s financial regulator has disclosed that MetaX has finalized the IPO preparation process known as regulatory tutoring, reports TechNode. Over a span of five years, MetaX completed eight investment rounds, securing several billion yuan in total capital. Its estimated market value reached 10 billion yuan (approximately $1.38 billion), according to the 2024 Hurun Unicorn Report.

Zhaoxin, a joint venture between the Shanghai government and Via Technologies, is better known for x86 CPUs with integrated graphics. The company recently applied to list on the Shanghai STAR Market, expected to be completed by mid-2025, which makes it one of the more advanced IPO candidates among China's semiconductor companies. A public offering could occur within the next 12 to 18 months if approvals proceed smoothly.

Biren Technology, Moore Threads, and MetaX are pursuing IPOs primarily to secure substantial funding for continued product development and scaling in a highly capital-intensive industry. All three companies are developing advanced AI and GPU hardware, which demands heavy investment in R&D, silicon validation, software ecosystems, and talent, especially as they aim to rival global leaders like Nvidia.

Beyond funding, public listings also serve strategic and reputational goals. IPOs enable these companies to demonstrate transparency, regulatory compliance, and long-term commitment, thereby positioning them as credible suppliers in China's push for self-reliance in semiconductors. For Moore Threads and MetaX in particular, an IPO can elevate their visibility among enterprise and cloud clients, while Biren may use the opportunity to cement its role as a national AI accelerator champion.

Some developers choose to stay private

The first traces of Moore Threads' GPU programming software stack, dubbed MUSA, have surfaced online, furthering the nation's pursuit of tech-autarky. MUSA serves as an alternative to Nvidia's CUDA environment, compatible with the domestic MUSA MTT GPU lineup. Any open-source pedigree of the SDK has not been mentioned, so it is likely proprietary and won't be of much benefit to developers outside China.

The U.S. has implemented a series of export restrictions on China, including: advanced AI chips, high-bandwidth memory (HBM), manufacturing equipment, and silicon wafers from leading players like Intel, TSMC, and Samsung. In a bid to reduce reliance on Western hardware, China is hard at work developing its semiconductor ecosystem with in-house silicon, fab equipment, memory, CPUs, and even GPUs. The latter is of great importance, as modern-day machine learning (sometimes under the buzzword banner of AI) is largely accelerated by parallel computing, something which GPUs excel at.

A strong GPU programming ecosystem offers high-level abstraction, ready-to-use libraries, documentation, and profiling tools. With high-performance Nvidia GPU exports still in limbo, Moore Threads is offering an alternative to CUDA.

MUSA provides a built-in compiler (MCC), runtime libraries (MUSA Runtime), a comprehensive list of specialized libraries (MUSA-X), debuggers, and profilers. To ensure compatibility with already written CUDA code, the MUSA SDK also includes Musify, a tool that translates CUDA code for the MUSA environment, likely by translating PTX code at runtime, similar to zLUDA.

The MUSA SDK version 4.0.1 is compatible with x86 processors from Intel (on Ubuntu) and Hygon (on Kylin). Moore Threads is demonstrating the prowess of its stack through several demonstrations on its website, including speech synthesis, AI-image generation, image processing, AI-powered 3D face modeling, just to name a few. You can actually try out a bunch of these demos right now (though you might need an account), some of which are reportedly running on Moore Threads' MTT S3000 datacenter GPUs.

Despite CUDA's clear advantage in terms of advancement, maturity, and support, MUSA could find many indigenous customers in small-scale environments, evolving over time. AI developers and researchers envision a heterogeneous future, championing the adoption of hardware-agnostic and open-source platforms. Breaking free from CUDA's reign requires superior alternatives, with ROCm being a key contender. However, AMD's hardware support still trails behind Nvidia.

Today, Chinese GPU manufacturer Moore Threads released new v290.100 drivers for its MTT S80 and S70 GPUs, boasting impressive performance improvements in select titles (via ITHome). The latest update addresses commonly reported bugs and stability issues across games and applications such as Unigine Valley, Rhinoceros 3D, and several local applications. The updated drivers are available for download on their website and support Windows 11/10 and Linux.

Driver v290.100 promises optimizations for A Plague Tale: Requiem, Death Stranding, Infinity Nikki, Super Power Continent, The King of Fighters XV, Sonic Forces, Steel Rats, Honkai: Star Rail and Genshin Impact. Bear with us, as a few of the aforementioned names may have been mistranslated.

Moore Threads claims a massive 120% increase in average FPS in A Plague Tale Requiem, followed by a 50% performance boost in Death Stranding. Let's be honest: these aren't mainstream titles anymore; some are incredibly niche. However, the two big shots from HoYoverse could draw some attention to these GPUs in the local market, though exact specifics or numbers haven't been shared.

The new drivers also fix several reported issues, including display anomalies in Rhino 8.11, crashes in Shadertoy, Rainbow Six: Siege, Devil May Cry 5, Unigine Valley, and several problems associated with the PES control center. However, a few unresolved bugs remain: unresponsive behavior in Blender and stability issues with local software.

Since launch, Moore Threads has been hard at work delivering driver optimizations for its GPUs. Truth be told, we aren't sure if we should take these numbers at face value. Despite these cards' raw horsepower, software, and certain architectural limitations are likely holding them back. Market share makes a huge difference, as it wouldn't make sense for developers to expend valuable resources to optimize their software for these GPUs.

Even Intel's Xe-based Alchemist suffered greatly from lackluster driver support and an immature architecture at launch. Battlemage mostly fixed these problems, with the Intel Arc B580 standing as a great budget GPU, rivaling Nvidia's RTX 4060 and even the potentially soon-to-launch RTX 5060. The silver lining is that due to the 16GB frame buffer, these GPUs reportedly show "excellent" performance in AI inference tasks. It's suggested that DeepSeek's R1 model runs inference not on Nvidia's Hopper/Blackwell chips but on Huawei's homegrown Ascend AI accelerators. Still, despite these recent strides, it might take years for Chinese hardware and software to match Western technology.

One of the breakthroughs of DeepSeek's open source AI models is that they can be run locally using relatively inexpensive hardware, like the Raspberry Pi.
As it turns out, the DeepSeek V3 and R1 models can even be run on Moore Threads GPUs developed in China, reports ITHome. If true, this is a major achievement for DeepSeek, the hardware designer, and China as this potentially opens new doors for Moore Threads and reduces reliance of DeepSeek and China on Nvidia hardware. 

Moore Threads reportedly says it had successfully deployed the DeepSeek-R1-Distill-Qwen-7B distilled model on its own MTT S80 client graphics card and MTT S4000 datacenter-grade graphics cards. The company used the Ollama lightweight framework that enables users to run large language models directly on their MacOS, Linux, and Windows machines as well as an optimized inference engine to achieve 'high' performance. 

Although the report claims 'excellent' and 'high' performance when describing how the MTT S80 and MTT S4000 performance with the DeepSeek-R1-Distill-Qwen-7B distilled model, it does not specify actual performance numbers or make comparisons to other hardware. To that end, it is impossible to evaluate the claims. Furthermore, given the fact that the MTT S80 is barely available outside of China, it is impossible to verify them. 

Ollama supports models like Llama 3.3, DeepSeek-R1, Phi-4, Mistral, and Gemma 2, enabling their efficient local execution without relying on cloud-based services. Ollama is developed primarily for macOS and uses Metal for Apple GPU acceleration, CUDA for Nvidia GPU acceleration, and ROCm for AMD GPU acceleration. 

Officially, Ollama does not support Moore Threads's GPUs, but the company claims that its graphics processors can execute code compiled for CUDA GPUs. The results confirmed that Moore Threads's GPUs are indeed compatible with CUDA and suitable for AI workloads, particularly in Chinese-language applications. 

To further enhance performance, Moore Threads employed a proprietary inference engine featuring custom computational optimizations and improved memory management. This software-hardware integration significantly boosts computing performance and resource efficiency and ensures smooth deployment process and supporting future AI models, according to the report. Of course, we are talking about a distilled model, so for now we cannot really compare performance of Moore Threads GPUs with performance of solutions from AMD, Apple, or Nvidia.

Chinese GPU designer Moore Threads has developed the MTT X300, a new graphics card for workstations. Under the company’s professional vision accelerator label, the X300 supports x86, Arm, and LoongArch CPUs. Moore Threads says the GPU works in Windows, Ubuntu, and Chinese OSes like Fangde and Tongxin.

On the hardware side, the X300 is a PCIe 5.0 card with 4,096 second-generation MUSA cores, 16GB of GDDR6 memory, and a 256-bit wide memory bus capable of hitting 448 GB/s of bandwidth. The X300 also has 8K video output, AV1 encoding and decoding, and support for common graphics APIs such as OpenGL, DirectX, and Vulkan.

However, this isn’t anything new for Moore Threads, as these specifications are identical to those of the MTT S80, launched over two years ago. The official spec sheets for both cards are nearly the same, with the only difference being that Moore Threads doesn’t specify the clock speed for the X300. Both cards are the same size, meaning they likely share the same cooler.

While there are seemingly no differences between the X300 and S80 concerning hardware, drivers are probably different. It’s pretty standard for Nvidia and AMD to use the same hardware for gaming and workstation GPUs (though not quite to this degree), but only owners of workstation cards get access to professional-grade drivers. These drivers, such as Nvidia's Studio drivers, come with specific optimizations for applications that aren’t games, like AutoCAD.

Moore Threads doesn’t explicitly say that the X300 has workstation-optimized drivers but does advertise that the card supports apps like Unreal Engine, Unity, and AutoCAD. Unfortunately, those drivers aren’t yet available on Moore Threads’ website, so we can’t know. At the very least, the X300 would be a pointless product if it didn’t have any special drivers since it would otherwise be identical to the S80.

How the X300 performs in workstation workloads is an open question. In games, even the Ryzen 5 8600G, with its integrated graphics, easily surpasses the S80, so the X300’s raw horsepower probably isn’t all that promising. It does have lots of decently fast memory, though, which could be a big selling point.

Moore Threads has also struggled with fully optimizing its GPU drivers, and just in October, it released new drivers that improved performance in two games by 30 to 40%. These sorts of driver updates are pretty typical for Moore Threads, indicating that lots of performance is potentially being left on the table, and that could be the case for professional applications.

A new report by the UK-China Transparency organization claims that British company Imagination Technologies has supplied Chinese companies with critical GPU IP technologies that could be used for China's government, military, and secret service applications.

In the report, UKCT accuses Imagination's managers and investor Canyon Bridge (which is controlled by China Reform, an arm of the Chinese government) of facilitating the transfer of key technology assets to Chinese companies, which now rival Imagination and other GPU developers. The technology was reportedly transferred to Moore Threads for example, which has links to a company that supplies GPUs to China's military, and Biren technology, which is partly owned by the Russian government. 

Imagination Technologies develops graphics and neural processing IP, which it then licenses to clients like Apple. Like some other major IP licensees, it can sell an actual GPU or NPU blueprint (physical IP), explain how to use it, and perhaps even help implement it into a chip design. It can also sell an architectural license, from which its partner would design a GPU from scratch — developing its own execution units (texture units, shader units, render back ends, etc.), data pipelines, cache hierarchies, and interconnects, as well as balancing the whole GPU in accordance with its vision.  

Here are two of the four key findings from the report, with further analysis below.

• Transferal of core assets to Chinese companies: UKCT has received testimony alleging that in 2020 Imagination began to transfer core assets to Chinese GPU companies through undisclosed and unusual knowledge and technology transfer deals. Of the three Chinese companies identified as having received Imagination’s assets: one, Moore Threads, has links to a company that supplies GPU products to the Chinese military; another, Biren Technology, is part-owned by the Russian government. These two companies have been described as China’s “premier AI chip designers” and both were sanctioned by the US government in October 2023.
• Chinese owners closely tied to military: Imagination’s owner, China Reform, focuses on supporting China’s strategic industries and has close links to China’s military and national security establishment as well as investment in related industries and companies, including a chip company that has supplied a leading military company. UKCT assesses that China Reform has intended to instrumentalise Imagination in pursuit of the interests of the Chinese Communist Party (CCP).

Licensing GPU technologies to China

Normally, Imagination and other IP vendors never transfer key knowledge or explain design decisions to customers. However, UKCT's source — a former employee of Imagination — alleges this occurred with Imagination's China-based customers, including Biren Technology (which has received investments from the Russia-China Investment Fund), InnoSilicon, and Moore Threads.

The UKCT source revealed that the architectural licenses sold to Chinese companies comprised three distinct components:

• The first included standard materials provided in Imagination's regular licensing deals, such as physical IPs, support, and documentation. These standard resources do not include details about the rationale or methods behind specific design decisions.
• The second involved a direct transfer of knowledge, where Imagination's senior hardware and software architects shared unique insights with Chinese clients. This was highly unusual and had only occurred once before — with Imagination's biggest customer, Apple.
• The third required the creation of new documentation specifically tailored to these Chinese clients, detailing foundational features of Imagination's design intellectual property (IP), including internal specifications, verification methods, and design processes. This kind of information is typically kept secret (though it is sometimes shared to meet compliance requirements, as it enables customers to become competitors).

The source felt that, after Imagination's know-how was transferred to three major GPU developers from China, the company would be shut down. The company did not shut down, but its staff count dropped from 1,100 in 2016 to 804 in mid-2017 (after Apple announced plans to develop its own GPUs) and further to 520 in 2023.

ImgTec denies wrongdoing

While InnoSilicon publicly acknowledged using Imagination's IP to develop its GPUs, Biren Technology and Moore Threads claim their processors rely on proprietary architectures. However, it is widely believed that the founding teams from Biren and Moore Threads came from AMD and Nvidia and potentially brought trade secrets with them. (This has not been confirmed, but neither the claims about proprietary architectures nor the assumption of third-party trade secrets preclude the possibility that the underlying instruction set architecture (ISA) or some design decisions originate from Imagination Technologies.)

Imagination Technologies, which is based in the U.K., denies any wrongdoing and claims that its close work with Chinese customers adhered to industry standards for architectural licenses. The company also denied transferring proprietary know-how to its licensing partners.

"Architectural license deals are entirely normal in the semiconductor industry, including for Imagination," a statement from Imagination Technologies published by UKCT reads. "When we 'supply IP,' we deliver technology packages and grant a license to Imagination's intellectual property rights in that technology to enable our customers to design, manufacture, and sell chips incorporating our technology. [...] The specific packages of technology deliverables for standard and architectural licenses vary by customer and end-use case [‚Ķ]. For standard and architectural licensing engagements, training is necessary to help customers understand how to design chips with Imagination Technology and, where relevant, to carry out their own customization for their permitted use only." 

The company further argued that it would not make strategic sense to share all of Imagination Technologies' experience and knowledge with customers, enabling them to design their own GPUs. It also denied transferring its assets to third parties in China. Now that Biren, InnoSilicon, and Moore Threads are on the U.S. Department of Commerce's Entity List and Imagination can no longer supply these companies, the Chinese firms likely have to develop new GPU generations independently — probably using a familiar ISA. 

Chinese Military Ties?

When China-backed Canyon Bridge took over Imagination Technologies in 2020, it immediately caused a huge controversy because of its ties to the Chinese government — and the potential use of its technologies to enhance China's military capability and strategic industries. UKCT asserts that Canyon Bridge not only has ties to the Chinese government through its owner, China Reform, but that it also has ties to the People's Liberation Army, state-owned strategic enterprises, and various Chinese companies that require AI, graphics, and HPC processors for development and production. These companies can collaborate with Biren, InnoSilicon, and Moore Threads — either directly or through intermediaries. 

For example, one of Moore Threads' shareholders is Liu Shanshan (also the company's executive director). Shanshan also serves as a director at Beijing Runyu Information Technology. According to UKCT, this GPU company has supplied products to China Aerospace Science and Technology Corp., a major arms manufacturer in which China reform has a stake. 

UKCT plans to release a second part of its investigation, which will focus on export control laws and produce further evidence from an employment dispute involving Imagination's former CEO, Dr. Ron Black. You can read the full report here [PDF].

Moore Threads, a Chinese GPU developer that is also focusing on AI, is progressing toward a potential IPO after restructuring as a joint-stock entity and increasing its capital base to ¥330 million ($45.6 million), reports the South China Morning Post. Despite being on the U.S. Entity List, the company has garnered substantial private and government funding to advance China's AI and GPU capabilities. 

In preparation for a public listing, Moore Threads shifted from a limited-liability company to a joint-stock status in late October, raising its capital from $3.32 million to $45.6 million. This restructuring, often a step toward an IPO, signals a likely move to enter the public market, which would align strategy of Moore Threads with other domestic players like Biren Technology and Enflame, also preparing for listings. 

Moore Threads was founded in 2020 by Zhang Jianzhong, who previously led Nvidia's China operations. So far, the company has introduced three generations of microarchitectures designed both for gaming and AI workloads. 

Moore Threads has benefited greatly from China's push toward technological independence as it got plenty of money from state-linked investors. Since its inception, the company has raised $800 million from notable investors, including ByteDance, Tencent, and state-backed Beijing Zhongguancun Science City Innovation Development, highlighting strong support from both private and public sectors.  

The company has positioned itself as a top GPU and AI-accelerator developer in China and became pretty well known, yet officially its focus was on gaming (probably to fly below the U.S. government's radar). Still, Moore Threads's second-generation Chunxiao architecture supports FP16 and INT8 data formats that are not required for graphics processing in addition to FP32, which is used for graphics. The company's 3rd Generation Quyuan GPU architecture already supports FP64, FP32, TF32, FP16, BF16, and INT8 data formats used for AI and HPC workloads and can scale to clusters containing up to 10,000 GPUs.  

As a result of the dual-use nature of Moore Threads GPUs, U.S. sanctions in October 2022 placed Moore Threads into the Entity List (along with Biren Technology, another China-based developer of compute GPUs), blocking it from using American technologies, including multinational foundries, such as TSMC, to produce chips. 

Following the Moore Threads DX12 enablement driver, the Chinese-based GPU manufacturer has released another driver in quick succession, this one featuring optimizations for new and existing titles and bug fixes. The driver, PES Control Center 270.80.2, features up to a 40% gaming performance improvement for its S80 and S70 graphics cards.

Moore Threads has introduced optimization improvements for five titles: Getting Darker, Predator, War Will, Sniper: Ghost Warrior Contact 2, and Man of War 2. Getting Darker sees the largest improvement, featuring a 40% performance improvement with driver 270.80.2. Likewise, Predator receives a 30% performance improvement. Moore Threads neglected to share any performance-related details for the last three titles, only stating War Will, Sniper: Ghost Warrior Contract 2, and Man of War 2 have received “game experience” optimizations.

The Chinese GPU manufacturer resolved two issues in this driver: Monster Hunter: World no longer suffers from problematic graphics anomalies after modifying image quality settings in the game menu. Video recordings will no longer flicker during the recording process, and the PES control center will no longer respond while recording. Driver 270.80.02 also introduces “full functionality” for “Mobi Ma Liang.”

It's great to see Moore Threads constantly improving its GPU drivers. Earlier this month, it finally introduced its first DX12-supported driver, enabling its gaming-focused MTT S-series GPUs to run DX12 titles. However, despite this, Moore Threads still suffers from significant performance issues, even on its latest DX12 driver, that make its flagship GPU run worse than a Turing-based GTX 1650.

As a result, Moore Threads will need to continue pushing driver updates such as the one covered today if it hopes to compete with more experienced manufacturers such as AMD and Nvidia in the future. Moore Threads is in a very similar state to Intel several years ago when it first began building GPU driver updates.

It took Intel several years of constant driver iterations to finetune its Arc Alchemist GPUs. The main difference with Moore Threads is that it is in an even worse state than Intel, with GPUs that can barely compete with entry-level cards from several years ago. Thus, it will need to rely more heavily on driver iterations to overcome this issue, assuming the hardware is not the bottleneck.

When Moore Threads introduced its MTT S80 and MTT S3000 graphics cards based on the Chunxiao architecture, roughly two years ago, the company touted both decent performance and feature sets. Yet, the company admitted that it would take some time to develop proper drivers supporting the latest application programming interfaces (APIs) and games. As noticed by @Loeschzwerg_3DC, it has taken the company about two years to release its first DirectX 12 driver.

Launching a beta driver supporting DirectX 12 is a rather big deal for Moore Threads. Its graphics boards have consistently lagged behind the best graphics cards from AMD and Nvidia both in terms of performance and features. For now, Moore Threads's MTT S80 can support numerous DirectX 12 capabilities (not all of them, of course, as the Chunxiao architecture does not support ray tracing, mesh shaders, VRS, etc.). Testing shows it can now run the 3DMark Time Spy benchmark, the Steel Nomad benchmark, and some Chinese games that require DX12.

Sadly, the performance of the Moore Thread MTT S80 graphics card does not really impress, according to preliminary tests conducted by @Loeschzwerg_3DC. The board scores 3,783 points in 3DMark Time Spy (3,452 graphics score, 8,315 CPU score), which is lower than the performance of Nvidia's GeForce GTX 1650 (around 4,400 points) when used with Intel's six-core Core i5-11400 CPU. Over time, Moore Threads can probably increase the performance of its MTT S80 when running DirectX 12 workloads, though this will unlikely make this graphics card competitive in terms of performance with modern add-in-boards from AMD, Intel, and Nvidia.

When tested with the DirectX API Overhead Feature Test (designed to test the overhead of DX11, DX12, and Mantle APIs), the MTT S80 scored 398,246 draw calls per second in multi-thread mode and 403,078 draw calls per second in single-thread mode. Such results are lower than those of Nvidia's GeForce GTX 1650 and look worse than those of 10-year old GPUs, which pretty much indicates the quality of Moore Threads's drivers in their present form.

Of course, we cannot draw solid conclusions about the quality of Moore Threads's DirectX 12 drivers based on a single report. However, considering that drivers have plagued the performance of the MTT S80 for two years, we really cannot expect the company's DX12 drivers to be perfect from day one.

Chinese GPU manufacturer Moore Threads has unveiled another S-series GPU in its consumer lineup. This time, it has debuted a mid-range model dubbed the S50, featuring 8GB of VRAM and 2,048 MUSA cores.

The S50 sits firmly in the middle of Moore Threads' consumer GPU product stack, flanked by the S60, S70, and S80 at the top and the S30 (4GB, 2GB) and S10 at the bottom. Specs-wise, the S50 is incredibly similar to the S60. The S50 features 2,048 cores, 8GB of memory, and a 256-bit memory bus, which are the exact same specifications as the S60. Both GPU variants also come packaged in a single-slot graphics card form factor.

Despite boasting identical specs, Moore Threads rates the S50 with less computing horsepower than the S60. The S50 is rated at 5.2 Teraflops of FP32 computing power and 20.8 TOPS of INT8 performance (a metric for AI-based workloads). In contrast, the S60 is rated with 15% more FP32 computing power, boasting 6 Teraflops of FP32 crunching power. For comparison, the S50's FP32 performance lands between the GTX 1070 and GTX 1060 6GB from eight years ago.

Frustratingly, Moore Threads does not list all the specs for the S60 or the S50 to allow us to find the main difference between the two. But we suspect it is down to power consumption and clock speeds. A large enough gap in power consumption and clock speeds would cause a 15% disparity in performance. The S50 comes with an 85W TGP, but we don't have a listed TGP spec for the S60.

Moore Threads advertises the S50 as having excellent 1080p performance in lightweight titles such as CS: GO and Dota 2 on Linux operating systems specifically. These unorthodox advertisements are probably a result of Moore Threads' fragile graphics drivers, which have been the Achilles heel of its mainstream graphics cards. 

Moore Threads' driver problem is similar to Intel's driver problems of the past but on steroids. For perspective, the company's S80 flagship struggled to compete with Nvidia's GT 1030 (that's not a typo) at its worst gaming performance a year ago. Recent driver updates this year have done little to rectify Moore Thread's massive disparity in gaming performance compared to Intel, AMD, and Nvidia's offerings.

The S50 also boasts support for all mainstream graphics APIs, such as DirectX, Vulkan, OpenGL, and OpenGL ES. Resolution support peaks at 8K, and the GPU has a single MUSA encoder and decoder that supports AV1, H.265 (HEVC), and H.264 video recording and playback.

The pricing for the S50 is unknown for now. The graphics card is currently available for free trial for corporate consumers. The S50 hasn't hit the retail market yet.

One of Nvidia's advantages in the data center space is that it not only offers leading-edge GPUs for AI and HPC computing but can also effectively scale the number of its processors across a data center using its own hardware and software. How could you defeat Nvidia if your GPUs are slower and your software stack is not as pervasive as Nvidia's CUDA? Well, expand your own scale-out capabilities. This is exactly what Chinese GPU maker Moores Threads has done, based on a Science China Morning Post report.

Moore Threads has upgraded its KUAE data center server for AI, enabling connecting up to 10,000 GPUs in a single cluster. The KUAE data center servers integrate eight MTT S4000 GPUs interconnected using the proprietary MTLink technology designed specifically for training and running large language models (LLMs). These GPUs are based on the MUSA architecture and feature 128 tensor cores and 48 GB GDDR6 memory with 768 GB/s of bandwidth. A 10,000-GPU cluster wields 1,280,000 tensor cores, but the actual performance is unknown as performance scaling depends on numerous factors.

This move highlights Moore Threads's efforts to boost its datacenter AI capabilities despite being on the U.S. Department of Commerce's Entity List. Moore Threads' products, of course, lag behind Nvidia's GPUs in terms of performance. Even Nvidia's A100 80 GB GPU introduced in 2020 offers compute performance significantly greater than that of the MTT S4000 (624/1248 INT8 TOPS vs 200 INT8 TOPS). Yet, there are claims that the MTT S4000 is competitive against unknown Nvidia GPUs.

Moore Threads, which was founded in 2020 by a former Nvidia China executive, does not have access to leading-edge process technologies due to U.S. export rules as it is blacklisted by the Biden administration. However, the company is developing new GPUs for gaming (these graphics cards aren't on our list of the best graphics cards) and is pushing forward in the AI sector despite significant obstacles.  

So far, Moore Threads has forged strategic partnerships with major state-run telecom operators, including China Mobile and China Unicom, as well as China Energy Engineering Corp. and Gulin Huajue Big Data Technology. These collaborations aim to develop three new computing cluster projects, further advancing China's AI capabilities. 

Moore Threads recently completed a financing round, raising up to 2.5 billion yuan (approximately US$343.7 million). This influx of funds is expected to support its ambitious expansion plans and technology advancements. However, without access to advanced process technologies offered by TSMC, Intel Foundry, and Samsung Foundry, the firm faces numerous challenges on the path to developing next-gen GPUs. 

Moore Threads claims to be making great strides in its AI GPU development, with its latest S4000 AI GPU accelerator being exponentially faster than its predecessor. As reported by cnBeta, a training regimen of a new Kua'e Qianka Intelligent Computing Cluster sporting S4000 GPUs ranked third fastest in AI testing, outperforming several counterparts consisting of Nvidia AI GPU clusters.

The benchmark run was taken from a stability test of the Kua'e Qianka Intelligent Computing Cluster. Training took a total of 13.2 days and supposedly ran perfectly with no faults or interruptions for the duration of the run. The AI model used to benchmark the new computing cluster was the MT-infini-3B large language model.

The new computer cluster reportedly ranks among the top AI GPU clusters of the same scale (using the same number of GPUs, presumably). However, the above table is decidedly lacking in details. The MTT S4000 cluster was compared to unspecified Nvidia GPUs, for example — we don't know if those are A100, H100, or H200 GPUs, but we suspect A100 to be most likely. The workloads are also not the same. Training MT-infini-3B could be quite different from training Llama3-3B for example. Sprinkle liberally with salt, in other words.

Even without being apples-to-apples, however, training LLMs on the Moore Threads GPUs represents an important step in China's domestic GPU roadmap. The Kua'e Qianka computing cluster at least suggests the MTT S4000 AI GPUs are competitive with Nvidia's prior-generation A100 architecture. This is backed up by the S4000's raw performance numbers, which not only significantly outperform Moore Thread's S3000 and S2000 AI GPU predecessors but also outperform Nvidia's Turing-based AI accelerators. The S4000 doesn't match Nvidia's A100 AI GPU accelerators, but perhaps it's not far away from Ampere performance levels.

For Moore Threads, the Kua'e Qianka's performance capability is a huge win regardless of what Nvidia GPUs or LLMs were tested. It demonstrates that Moore Threads is now capable of building AI GPUs that can perform similar work as AI GPU competitors from Nvidia, AMD, and Intel. It might not perform better, but it's an important stepping stone on the path to faster and more capable supercomputers and AI clusters.

It's an impressive feat for a GPU manufacturer that was founded less than five years ago. If Moore Threads can keep delivering significant generational performance improvements, it could have an AI GPU accelerator that can go toe-to-toe with its Western counterparts in the next few years. That's a big "if" of course, and we know from historical precedent that GPU development doesn't always go as planned.

We'll also be interested to see if Moore Threads can put its seemingly good AI performance capabilities into its gaming graphics cards. To date, the MTT GPUs have suffered badly in gaming tests, thanks in part to immature drivers/optimizations. While AI needs lots of computational power, it's different from real-time computer graphics, so expertise in one area doesn't imply similar capabilities in the other.

The city government in Beijing has outlined plans to subsidize companies that buy Chinese-made processors, according to a report published by Hong Kong’s SCMP. The draft policy has a strong emphasis on GPUs as they have been most badly affected by US export controls, and China desperately needs GPUs to advance its AI initiatives.

The Beijing Municipal Bureau of Economy and Information Technology will provide undisclosed subsidies towards the purchase of domestically made GPUs to “expedite the supply of controllable intelligent computing resources,” says a draft policy document. The aim, perhaps optimistically, is that Beijing will achieve “self-sufficiency” with regard to computing infrastructure by 2027.

We have reported on and looked closely at various ‘homegrown’ Chinese GPUs over recent years. In December 2022 we summarized the competitive landscape of the GPU market worldwide (18 active firms), noting that China had rapidly boosted the numbers of domestic GPU developers. Typically, the pre-launch claims of high performance and efficiency coming from these Chinese firms don’t quite live up to expectations post-launch. This also happens in the West, but firms like AMD and Nvidia are already in a good position, making some of the best graphics cards for consumers.

We have seen recent reports of China getting its hands on sanctioned technologies and products like the Nvidia H100 through indirect means. However, such supply routes aren’t usually sustainable or reliable enough for serious business, and China is thus determined to make the most of its domestic resources to fix this issue.

The Chinese GPU makers best placed to benefit from the GPU and AI processing subsidies may be firms like Biren Technology, Moore Threads, Innosilicon (PowerVR), MetaX, and Zhaoxin (Glenfly Arise). Loongson also recently debuted a homegrown GPU which is said to be comparable to the AMD Radeon RX 550 in terms of FP32 performance. Perhaps some other Chinese GPU firms are in gestation, or ready to hatch from Chinese universities and research institutes.

Beijing is said to be home to more than half of the LLMs that have been developed in China. According to the SCMP report, the selected GPUs and subsidized projects will be largely tasked with training general and industry-specific large language models (LLMs). Meanwhile, R&D in complementary technologies such as software, dedicated AI processors, silicon photonics, and quantum computing chips also continues apace.

An open source fully custom GPU has come out of stealth after four years in development. FuryGPU has been a one-man effort from games software developer Dylan Barrie, who says he put together this extremely complex hardware and software project in his spare time. FuryGPU is based on a Xilinx FPGA design, and the prototype PCIe graphics card is currently capable of achieving around 44fps in the Quake Timedemo. The work on FuryGPU was undertaken after Barrie was inspired by Ben Eater’s building a programmable 8-bit computer from scratch project.

As you can see from the pictures in this article, FuryGPU looks very much like a typical PC graphics card from about 20 years ago, modernized by equipping DisplayPort and HDMI outputs. The project is far more than the hardware, though, with Barrie admitting that the most painful aspect of this graphics card’s design was creating the Windows drivers.

Hardware, from maker board to graphics card

Barrie began realizing his dreams of building a GPU from scratch after picking up an FPGA-packing Arty Z7 development board and doing some preliminary development and testing. Subsequently, the project got a boost by the debut of Xilinx Kria System-on-Modules (SoMs), which combine “insanely cheap Zynq UltraScale+ FPGAs with a ton of DSP units and a (comparatively) massive amount of LUTs and FFs, and of particular interest, a hardened PCIe core,” enthused Barrie.

To go from this maker board to the FuryGPU PCIe add-in-card design we see in 2024, Barrie taught himself the SystemVerilog hardware description and hardware verification language and the KiCAD EDA / electronics CAD software suite. He says a Herculean effort was required to design the schematic for the FuryGPU with the 4-lane PCIe we see today, even with the FPGA circuitry built into the SoM. Now it was time to plug the FuryGPU into his test rig, write the drivers, and test games.

Windows drivers and Quake at 60fps

Creating the Windows drivers for FuryGPU is described by Barrie as “the most painful” aspect of the entire project – despite his day job being in the software side of graphics rendering in the games development industry for the last 14 years.

Initially, the FuryGPU maker’s ambition was to put together a simple spinning cube demo, to show the GPU working. However, as the project developed, playing the iconic PC game Quake at playable frame rates started to become the new goal.

Barrie explains that after getting the Windows drivers ready, he wrote a custom graphics API to communicate with the GPU, wrote Windows kernel drivers for the display and audio, and now has “a fully-functional piece of graphics hardware that can render Quake at a solid 60 frames per second.”

We’ve embedded Barrie’s Quake Timedemo video capture, demonstrating that the FuryGPU could achieve 44fps in this benchmark at 720p about a month ago. The developer says there are clear opportunities to get Quake “running much faster,” as he saw some obvious bottlenecks that he will target for optimization efforts.

The FuryGPU is set to be open-sourced. “I am intending on open-sourcing the entire stack (PCB schematic/layout, all the HDL, Windows WDDM drivers, API runtime drivers, and Quake ported to use the API) at some point, but there are a number of legal issues,” Barrie wrote in a Hacker News post on Wednesday. Because he works in a tangentially related vocation, he wants to make sure none of this work would break his work contract or licensing etc. This same thread includes quite a bit of extra detail about the FuryGPU project for those particularly interested.

On the FuryGPU website, there is an article dedicated to the GPU’s texture units, for those wishing to take a deeper dive into the architecture.

To conclude our coverage of this interesting new (to us) project, it is worth explaining the intended scope of the FuryGPU project. It is clear this is a maker project, like the breadboard CPU, but the FuryGPU offers such impressive performance that some may mistake it for a serious new GPU architecture. Barrie makes it clear this is not the case in the afore-linked Hacker News thread (writing under the nom de plume of PfhorSlayer). “This is a toy,” asserted the FuryGPU maker. "This is not going to change the GPU landscape or compete with any of the commercial players.” 

Even though the FuryGPU (or its offspring) may never make our chart of the best graphics cards, we will watch FuryGPU developments with great interest. Now that the project has gone public, there’s a chance that publicity and expert collaborators can accelerate the plans already in place.

Ryzen 8000G APUs continue to demonstrate why they are some of the best CPUs for gaming (specifically the Ryzen 5 8600G is our entry level alternative gaming pick). According to the German publication Computerbase, AMD's desktop Zen 4 chips with on-chip RDNA 3 graphics outperformed the Moore Threads MTT S80 and MTT S30, two of China's fastest domestic gaming graphics cards.

Like Intel's first-generation of Arc Alchemist graphics cards, Moore Threads also faced many performance and compatibility issues with its MTT graphics cards. The hardware wasn't holding back the MTT S80 and MTT S30, but rather the drivers. The drivers have gradually improved the MTT S80 and MTT S30 performance, which drove Computerbase to benchmark Moore Threads' graphics cards to see how they stack up against comparable Nvidia and AMD rivals.

The Ryzen 7 8700G was up to 90% faster than the MTT S80. With overclocked DDR5 memory, the Ryzen 7 8700G pushed the margin to 112%. The Ryzen 5 8600G was also faster than the MTT S80, posting 60% higher gaming performance. The MTT S80 was marginally better than AMD's integrated Vega graphics. The MTT S80 achieved 8% higher performance than the previous Ryzen 7 5700G, which wields Zen 3 execution cores with Radeon Vega graphics.

The MTT S30 was an utter disappointment, though. Since it targets office use cases, the level of performance was expected, and it's essentially an alternative to integrated graphics. The MTT S30 couldn't even outperform the Intel UHD Graphics 770 engine integrated into the Core i5-12500.

Computerbase tested the different graphics cards and APUs at a 1920 x 1080 resolution. The outlet specifically paired the MTT S80 and MTT S30 with a Ryzen 7 5800X and 16GB of DDR4-3200 C14 memory on the MSI MAG B550 Tomahawk. It used the latest 240.60 drivers.

Meanwhile, the Core i5-12500 and Ryzen 5 7500F powered the GeForce GTX 1650 and Radeon RX 6400 systems, respectively. As for the Ryzen 8000G tests, Computerbase utilized DDR5-5200 C32 memory for the base tests and faster DDR5-7200 C34 memory for the overclocked tests.

Although the new drivers have improved the MTT S80 and MTT S30 performance, much remains to be done. The China-made graphics card only works with some DirectX 11 games, and the DirectX 12 ones are out of the picture. The MTT S80 and MTT S30 were most effective in popular online multiplayer titles, which is likely the graphics card's target audience. The MTT S80 was decent in Dota 2 and Counter-Strike 2, offering average frame rates of 66.8 FPS and 59.5 FPS, respectively. 

Moore Threads is already preparing the MTT S90 and recently laid off some of its workforce due to U.S. sanctions. So it's unknown if the company has altered the resources dedicated to driver development. Either way, we should still see driver enhancements since the MTT S90 is unlikely to feature a new architecture.

Moore Threads, potentially China's top domestic GPU maker, has introduced its office productivity focused MTT S30 graphics card (H/T to VideoCardz). If you were underwhelmed by the gaming-centric MTT S80 and S70, prepare to be even less-whelmed by this significantly lower power compact graphics card. The various Moore Threads GPUs, including the MTT S30, are highly unlikely to ever make it into our best graphics cards roundup.

The specs you of the Moore Threads MTT S30 are already known. However, the performance of this family of GPUs is rather variable. We think this is due to driver immaturity, as huge gains are still being made with new driver revisions. Many bug fixes are also listed in the release notes.

New games supported in the latest driver include: Baldur’s Gate 3, Crossfire, World of Tanks, and Crysis 3. Unfortunately, some games have been accidentally broken by prior driver updates. Hopefully, the developers haven’t included any regressions with this new driver. Games won’t be of interest to MTT S30 users though.

According to the source, support for the Moore Threads MTT S30 is one of the new features of driver version 260.60. While the MTT S30 was announced several months ago, it hasn’t become readily available, but the driver is expected to signal the beginning of a full rollout.

Looking at the specs, it's obvious that the Moore Threads MTT S10 and MTT S30 are only for those who need a basic graphics adapter for office work. These cards may have an advantage in basic business environments with their low power consumption and no requirement for external power. Moreover, you can see that they can be used with the existing VGA monitor stocks which will be found in some offices.

There are also rumors (via VideoCardz) that an MTT S50 solution is on the way, with 2,048 Stream Processors, 8GB of VRAM, and a 75W TGP, to be marketed at workstations. Keep in mind that, at present, the fastest MTT S80 only competes with GTX 1650 hardware, at least in gaming performance. That's despite having triple the power budget and theoretically five times the compute, so we wouldn't expect an S50 to actually be all that performant.

One final item of interest is that the Moore Threads MTT S30 graphics cards recently acquired by the VGA Museum use a slightly different design to the pre-release models seen in photographs. Whether that affects performance or just represents a minor revision to clean up power or other aspects isn't clear.

China Telecom claims it has built the country's first supercomputer constructed entirely with Chinese-made components and technology (via ITHome). Based in Wuhan, the Central Intelligent Computing Center supercomputer is reportedly built for AI and can train large language models (LLM) with trillions of parameters. Although China has built supercomputers with domestic hardware and software before, going entirely domestic is a new milestone for the country's tech industry.

Exact details on the Central Intelligent Computing Center are scarce. What's clear so far: The supercomputer is purportedly made with only Chinese parts; it can train AI models with trillions of parameters; and it uses liquid cooling. It's unclear exactly how much performance the supercomputer has. A five-exaflop figure is mentioned in ITHome's report, but to our eyes it seems that the publication was talking about the total computational power of China Telecom's supercomputers, and not just this one.

We probably can't expect official performance benchmarks any time soon either, as China is neglecting to submit its supercomputers to TOP500, the organization that tracks the 500 fastest supercomputers in the world. This caginess is apparently down to fears about getting too much attention and inviting even more U.S. sanctions. 

It's hard to guess what might be inside this supercomputer, given the lack of details. On the CPU side of things, it may use Zhaoxin's KaiSheng KH-40000 server CPUs, which are now available in domestically-made servers. There are also other candidates though, like Loongson's 32-core 3D5000 and Phytium's 64-core Feiteng Tengyun S2500. All three chips differ greatly in respect to architecture, with Zhaoxin using x86 like Intel and AMD. Loongson uses a derivative of MIPS, and Phytium runs Arm's architecture.

Similarly, there are plenty of options for Chinese-made GPUs, with possibilities ranging from Moore Threads, Loongson, and Biren. Of the three companies, Moore Threads is the most recent to launch a new GPU in the form of its MTT S4000, which is already planned to see use in the KUAE Intelligent Computing Center. Loongson's LG200 arrived about two weeks before the S4000, though its claimed performance would make for a very slow supercomputer. Biren's BR100 would be a heavyweight champion, but it's unclear if it ever returned to production anywhere after TSMC stopped making it due to U.S. sanctions.

Regardless of the actual hardware inside China Telecom's new supercomputer, that it is reportedly made from top to bottom with Chinese hardware is the most important part. Relying solely on Chinese technology likely means the Central Intelligent Computing Center is disadvantaged in some or many areas. But technological independence is a key goal for China, even if it means swapping out cutting-edge Western hardware for slower but natively-made components. U.S. sanctions won't have much of an impact if China can manage to do everything itself.

Moore Threads has published a new driver for its MTT S70 and S80 gaming graphics cards, which aren't rivals for the best graphics cards yet. The new driver reportedly delivers significant performance gains in four games, such as 20% in Grand Theft Auto V. Although this driver update doesn't provide quite the same improvement as the company has in the past, it's still a significant boost for China's best domestic gaming GPU.

Grand Theft Auto V is the star of the new driver with its 20% gain in performance, but ~10% improvements were also made in three other popular titles: Lost Ark, Valorant, and Assetto Corsa. Although the MTT S80 could barely compete a year ago with even cards from the early 2010s, successive driver updates may put the card at least on par with today's lowest-end GPUs, such as the RX 6400 and Arc A380.

There were also swathes of bug fixes in the driver, tackling a problem that has plagued Moore Threads's gaming graphics since launch day. The new driver has 34 bug fixes, resolving crashes, blue screens, rendering issues, and unresponsiveness. By contrast, AMD only had seven bugs in its latest driver, and Nvidia had only three. Although many bug fixes are a sign that there are many bugs to fix, it seems Moore Threads is resolving them diligently, as it releases about one new driver a month.

While hardware is undoubtedly vital in gaming hardware, drivers are also a crucial component. Intel is another newcomer to gaming graphics cards, and it has had to issue tons of driver updates to get its Arc GPUs up to scratch. It's common to see Intel doubling or even tripling performance in specific titles with new drivers, as the company did with its latest driver on Thursday. It's probably safe to say that similar performance boosts are likely for future Moore Threads drivers, though given Moore Threads's smaller size, it might not be able to match Intel's pace.

Moore Threads has lately seen hard times, with U.S. sanctions on its graphics cards potentially causing a round of layoffs at the Chinese GPU designer. The company has maintained a steady pace of one monthly driver update since the layoffs. Still, the performance gains in this month's and last month's driver updates pale compared to November's, which saw much more significant gains and added support for over a dozen titles.

Chinese GPU manufacturer Moore Threads announced the MTT S4000, its latest graphics card for AI and data center compute workloads. The company's brand-new flagship will feature in the KUAE Intelligent Computing Center, a data center containing clusters of 1,000 S4000 GPUs each. Moore Threads is also partnering with many other Chinese companies, including Lenovo, to get its KUAE hardware and software ecosystem off the ground.

Although Moore Threads didn't reveal everything there is to know about its S4000 GPU, it's certainly a major improvement over the S2000 and S3000. Compared to the S2000, the S4000 has over twice the FP32 performance, five times the INT8 performance, 50% more VRAM, and presumably lots more memory bandwidth too. The new flagship also makes use of the second generation MUSA (Moore Threads Unified System Architecture) architecture, while the S2000/S3000 used the first generation architecture.

(Disclaimer: Moore Threads lists both the S2000 and S3000 as "First Generation MUSA," but others have said the S2000 and S80 used a "second generation Chunxiao architecture." It would make more sense for S3000 to be "second generation," as Moore Threads specifically calls the S4000 "third generation," though we don't have a product page for it yet.)

Compared to models from Nvidia, the S4000 is better than the Turing-based Tesla server GPUs from 2018 but still behind Ampere and Ada Lovelace, which launched in 2020 and 2022, respectively. The S4000 is especially lacking in raw horsepower, but it still has quite a bit of memory capacity and bandwidth, which may come in handy for the AI and large language model (LLM) workloads Moore Threads envisions its flagship will be used for.

The S4000 also has critical GPU-to-GPU data capabilities, with a 240 GB/s data link from one card to another and RDMA support. This is a far cry from NVLink's 900 GB/s bandwidth on Hopper, but the S4000 is presumably a much weaker GPU, making such a high amount of bandwidth overkill.

Alongside the S4000, Moore Threads also revealed its KUAE Intelligent Computing Center. The company describes it as a "full-stack solution integrating software and hardware," with the full-featured S4000 GPU as the centerpiece. KUAE clusters use MCCX D800 GPU servers, which each have eight S4000 cards. Moore Threads says each KUAE Kilocard Cluster has 1,000 GPUs, which means a total of 125 MCCX D800 servers per cluster.

On the software side, Moore Threads claims KUAE supports mainstream large language models like GPT and frameworks like DeepSpeed. The company's MUSIFY tool apparently allows the S4000 to work with the CUDA software ecosystem based on Nvidia GPUs, which saves Moore Threads and China's software industry from having to reinvent the wheel.

A KUAE cluster can apparently train an AI model in roughly a month, though it greatly depends on some particulars. For instance, Moore Threads says the Aquila2 model with 70 billion parameters takes 33 days to train, but bumping the parameters up to 130 billion will increase the time to train to 56 days.

Supporting this kind of hardware and software ecosystem would be challenging for any company, but it would be nearly impossible for Moore Threads to go it alone, especially after it had to lay off many of its employees. That's presumably why the company has established the Intelligent Computing and Large Model Ecological Alliance, a partnership between Moore Threads and several other Chinese companies. The domestic Chinese GPU manufacturer most notably got support from Lenovo, which has an international presence as well.

Although Moore Threads certainly won't be going toe-to-toe with the likes of Nvidia, AMD, or Intel any time soon, that isn't entirely necessary for China. US sanctions have prevented the export of powerful GPUs to China, which has not only given China's native semiconductor industry a good reason to exist, but also weakens the competition for companies like Moore Threads and its competitor Biren. Compared to Nvidia's China-specific cards, the S4000 and KUAE might have good odds.

Nvidia's GeForce RTX 4090D— the company's gaming graphics card designed specifically for Chinese market — is on track and its specifications are getting shape, according to recent leaks. Given the fact that the product has never been confirmed by Nvidia as well as the stance of the U.S. Department of Commerce regarding powerful GPUs for China, we can never be sure either about the specs or about the launch of the product at all. But TGP and clock leaks from @Zed__Wang and BenchLife.info we have some information. As ever, take leaks with a pinch of salt.

Nvidia's GeForce RTX 4090D is based on the AD102-250 GPU, and likely to have a base clock of 2280 MHz (up from 2230 MHz on the regular RTX 4090D) and posible boost clock of 2520 MHz (same as on the regular RTX 4090), according to @Zed__Wang, a renowned leaker. These clocks heavily suggest that the AD102-250 graphics processing unit will come with a reduced number of CUDA cores and other units, though we are speculating here. 

Now, with similar clocks and reduced number of CUDA cores, a GeForce RTX 4090D could match the performance of an RTX 4090 if overclocked (remember that the AD102 GPU was designed to be overclockable). So, to avoid this, the graphics card will come with total graphics power (TGP) capped at 425W (down from 450W in case of the original RTX 4090) and its overclocking capabilities will be locked, according to BenchLife.info.

The website reports that that Nvidia's add-in-board (AIB) partners are expected to receive samples of the AD102-250 GPU for testing this week, which means that we are going to see leaked specifications of the GeForce RTX 4090D shortly.

Now, while it is plausible that Nvidia is interested in making its $1599 graphics card available in China again,  and is willing to cut-down the performance of its AD102 below the threshold set by the U.S. Department of Commerce. It should be noted that the latter is unwilling to let Chinese have powerful GPUs that could be used for AI training or even technical computing at all. To that end, take any information about Nvidia's RTX 4090D or datacenter-oriented GPUs specifically tailored for China with a large grain of salt as the U.S. DoC could issue export rules and ban exports of these products to China even before their formal launch.

As we reported yesterday, Nvidia is reportedly building a version of its GeForce RTX 4090 product — presumably called the GeForce RTX 4090D. This version will comply with the latest U.S. export rules concerning hardware that could be used for artificial intelligence. Today, renowned hardware leaker @Zed_Wang revealed that the graphics cards will allegedly be based on the AD102-250 graphics processor. As this is a leak, take the news with some salt.

The GeForce RTX 4090D could theoretically make our list of the best graphics cards, if it's not limited to China. While the specs are unknown, we can hazard a guess that it will outperform the GeForce RTX 4080 but not reach RTX 4090 levels of performance. The RTX 4090 carries the AD102-300/AD102-301 GPU. The RTX 4090D looks set to be based on a significantly cut-down AD102-250 processor. At this point it's hard to say how the GeForce RTX 4090D will stack up against possible GeForce RTX 4080 Ti or GeForce RTX 4080 Super products, but it will likely sit higher up the GPU benchmarks hierarchy. It will also likely cost quite a bit more that a 4080 refresh.

Nvidia needs to lower the FP8 and FP16 performance of the GeForce RTX 4090 by at least 10% in order to reduce its AI potential, and decrease its TPP score from 5,286 to lower than 4,800 points. Nvidia cannot just cut down the Tensor core FP8 and FP16 processing, as the number of Tensor cores is tied to the number of SMs (Streaming Multiprocessors). How exactly Nvidia will choose to configure its AD102-250 to comply with the U.S. export rules is currently unknown, though there are multiple possibilities.

On the one hand, the company could reduce the number of SMs, CUDA cores, and Tensor cores on the GeForce RTX 4090D, which would increase product yields and cut costs. However, GPUs with a reduced number of active units tend to be good overclockers, so any 'lost' AI performance could be 'restored' in that fashion. Nvidia would either have to significantly cut the number of CUDA cores and risk that factory-overclocked GeForce RTX 4080 Super/RTX 4080 Ti would overlap with these parts, or moderately cut down the number of CUDA cores and prohibit heavy overclocking.

Besides the exact specifications, we also don't know precisely when the RTX 4090D is set to emerge. The earlier the better, of course, but the company will have to test everything before shipping to play it safe with the U.S. government. We suspect Nvidia will likely release the new device well before February 10, the Chinese New Year, as factories tend to shut down for a couple of weeks to celebrate.

Moore Threads is China's only developer of gaming graphics cards, and its first generation MTT S80 GPU just got a big driver update (via ITHome) that boosts performance tremendously, an important development now that the US government has banned Nvidia's fastest gaming GPU, the RTX 4090, from the country. The MTT S80's performance has now increased by double digits in five AAA titles, with one game getting an 80% improvement. For first-generation products, drivers are almost as important as the hardware itself, and it looks like the MTT S80 has room to grow even a year after it was released.

The new 240.50 version driver delivered the largest gains in PlayerUnknown's Battlegrounds, with an 80% boost to the average framerate. Meanwhile, the weakest improvement of the five was in Far Cry: New Dawn, which still saw an improvement of 30%. Valorant and Grand Theft Auto V both stood at 60% more frames, and Assetto Corsa registered a 50% bump.

Additionally, the new driver also added support for 18 other games. These games may have actually run on the MTT S80 before, but perhaps not very well. Moore Threads didn't detail how much better these games now run on the S80, though, which implies either their driver-side optimization didn't have much of an impact, or that the games were so unplayable before that putting a number down would make little sense.

Drivers have always been tough to nail down for GPUs sporting new architectures, especially when those GPUs are also the company's very first product. Even Intel, which had years of experience making drivers for integrated GPUs, has struggled to fully optimize its Arc Alchemist GPUs. Having launched over a year ago, Arc graphics cards are still getting driver updates with double and even triple-digit performance boosts.

Moore Threads has more problems beyond supporting a first-generation product. The company recently laid off a substantial amount of its workforce, which wasn't all that big to begin with. This is all thanks to U.S. sanctions against Chinese GPUs, such as the ones that Moore Threads makes. Sanctions apparently haven't stopped the company from working on its next-generation MTT S90 gaming GPU, which will hopefully benefit from the S80's maturing drivers. 

This latest round of drivers signals that the company is intent upon a steady cadence of updates, as its last round of driver updates earlier this year also provided substantial improvements. 

It's been less than a month since the U.S. government issued new sanctions against Chinese GPUs, but they're already having an effect, as Moore Threads is announcing layoffs (via TrendForce). The Chinese company, which is unique in that it has already launched GPUs for gaming and servers, is also restructuring its business by creating two new groups, one for AI and one for the metaverse.

It's not known how many employees will be laid off, nor is it clear how many employees Moore Threads even had prior to the layoffs. However, according to the company's LinkedIn page, there are between 500 and 1000 employees, which is already much fewer than GPU designing firms like AMD and Nvidia, both of which have roughly 26,000 employees (though many of AMD's people are CPU-centric rather than focused on GPUs).

Moore Threads' CEO, Jianzhong Zhang, said that the layoffs haven't impacted development of further GPUs and that there's still a path to success for the company. Notably, Zhang was Nvidia's China General Manager, who left the company and went on to found Moore Threads.

Alongside a slimming down of the workforce, Moore Threads is also reportedly doing some serious restructuring. It will create two new entities: The AI Strategy Group and the Metaverse Computing Strategy Group.

Ostensibly, Moore Threads sees lots of promise in those two fields, even though development of artificial intelligence (or AI) in China was what prompted the American sanctions in the first place. But given the exploding demand for hardware, and sanctions keeping Nvidia out of major markets, there's clearly room for another player if the company can continue to improve its products.

New driver optimizations have enabled Chinese vendor Moore Threads' MTT S80 gaming graphics card to rival Nvidia's GeForce GTX 1650 at 4K gaming. There may be some untapped performance, so don't be surprised if the MTT S80 eventually becomes a contender for the best graphics cards.

Armed with 4,096 MUSA (Moore Threads Unified System Architecture) cores with a 1.8 GHz boost clock, the MTT S80 is a PCIe 5.0 graphics card that pumps 14.4 TFLOPs of FP32 performance—the 7nm graphics card sports also sports 128 Tensor cores and 16GB of 14 Gbps GDDR6 memory. With access to a 256-bit memory interface, the MTT S80, which has a 255W TDP, delivers a memory bandwidth of up to 448 GB/s. The specifications look very decent on paper, minus the high TDP. However, the driver is what's hindering MTT S80's performance. The MTT S80 reportedly leverages the Chunxiao silicon, based on the PowerVR architecture developed by Imagination Technologies. The exact generation of the PowerVR GPU remains a mystery. Nonetheless, there's a lot of optimization work to get the graphics card to play nice with DirectX 11 titles.

Early benchmarks of the MTT S80 weren't very compelling. The Chinese homebrew graphics card lagged behind archaic mainstream performers, such as the GT 1030 in DirectX 9 games and GTX 1050 Ti in DirectX 11 games. Moore Threads recently deployed a new driver update, which the company claims to improve the MTT S80's performance by up to 40% in some titles. Chinese news outlet Expreview has put the MTT S80 through its paces with the latest driver (230.40.0.2), and the progress is pretty remarkable if you look at it from an unbiased standpoint.

Expreview's testbed consists of a Core i7-13700K processor, an Asus ROG Maximus Z790 Dark Hero motherboard, and 32GB (2x16GB) of G.Skill Trident Z5 RGB DDR5-7200 C34 memory on a system with Windows 10 21H2 64 bit operating system. The GeForce GTX 1650 used in the comparison is the GeForce GTX 1650 XC from EVGA, but Expreview didn't specify whether it was the Black Gaming or Ultra Gaming variant. The publication used a mixed bag of synthetic and real-world testing, but we'll only concentrate on the latter, which is more meaningful for gamers.

The overview is that the GeForce GTX 1650 was faster at 1080p (1920x1080) and 2K (2560x1440) resolutions, while the MTT S80 excelled at 4K (3840x2160). It makes sense since higher resolutions are more demanding on the VRAM requirement, giving the MTT S80 the upper hand. The MTT S80 has 4X more memory than the GeForce GTX 1650 and almost 3.5X higher memory bandwidth. The GeForce GTX 1650 is way more power-efficient compared to the MTT S80.

In Final Fantasy XIV, the GeForce GTX 1650 outperformed the MTT S80 by 37% in 1080p and 9% in 2K. However, the MTT S80 got its revenge at 4K, beating the GeForce GTX 1650 by 15%. The GeForce GTX 1650 delivered 49% and 45% higher frame rates at 1080p and 2K, respectively, in League of Legends. Meanwhile, the MTT S80 pumped 15% better frame rates at 4K.

We saw similar margins with Valorant. The GeForce GTX 1650 exhibited an 80% lead at 1080p and 17% at 2K. Conversely, the MTT S80 was 27% better at 4K. Finally, the performance margin was 53% at 1080p and 24% at 2K in favor of the GeForce GTX 1650 in Assetto Corsa. The MTT S80 led at 4K with a 9% performance delta.

Moore Threads MTT S80 Benchmarks

The MTT S80 is another example of how a mediocre driver can hold back a good product. It happens even to the more prominent manufacturers, such as Intel and its Arc Alchemist graphics cards. The chipmaker constantly improves performance with every new driver update, with the latest one claiming up to a 119% uplift. 

It's been a year since the MTT S80 hit the Chinese retail market. The 7nm gaming graphics card has gone through a lot over the last year. Moore Threads has released 12 driver updates, nine in the critical category. The software engineers at Moore Threads aren't miracle workers, but their efforts haven't been unsuccessful. The improvement from the May driver (221.31) and October driver (230.40.0.2) was up to 45% in some games.

Theoretically, the MTT S80 could compete with the GeForce RTX 3060. However, the Chinese graphics card still has a long way to go since it can't even consistently beat the GeForce GTX 1650, and there's a sizeable gap between the GeForce GTX 1650 and GeForce RTX 3060. The MTT S80 has recently dropped to $164, but it's still more expensive than the GeForce GTX 1650, which starts at $150 in the Chinese market. While we don't expect the GeForce GTX 1650 performance to increase anymore, the biggest issue with the MTT S80 is that it's unknown how much performance is left in the tank.

The Moore Threads MTT S80 graphics card has been subject to a significant price cut in its home country (h/t VideoCardz) in honor of Singles' Day. However, even at the new price of ¥1,199 ($164), it is not likely to fly off the shelves because issues like real-world performance and idle power consumption weigh heavily against it.

On paper the MTT S80 looks pretty attractive: it packs a modern 12nm PowerVR-based GPU with 4096 GPU cores, 16GB of GDDR6, and support for the latest PCIe Gen5 spec. Moore’s Threads claims a peak performance of 14.2 TFLOPS, and the memory is said to deliver 448 GB/s bandwidth.

But the MTT S80 excellent specs are less appealing when we start to look at the software / driver side of things. Support for new PC games and apps is limited, for example. This GPU is advertised as DirectX 11 capable, but it still seems happier with DX 9 titles.

Things look even worse for the Chinese graphics card when put under the beady eyes of testers outside of China. We have reported on both BullsLab Jay testing the MTT S80 (back in Feb.) and, more recently, Japan’s PC Watch checking out the card through an extensive test suite (in June).

When tested, the evidence shows that the MTT S80 is something of a sheep in wolf’s clothing. It has good specs on paper and even looks the part with its beefy triple fan cooling shroud. However, tests show the 255W MTT S80 performs somewhat similarly to an old 30W Nvidia GeForce GT 1030. Meanwhile, a modest GPU such as the GTX 1050 Ti absolutely towers over the MTT S80 — and has much better support in APIs, apps, and games.

People might be attracted by the MTT S80 (and S70) graphics cards at their ‘limited time’ pricing for 11.11 Singles' Day. But if they care to do the teeniest bit of research,  Chinese PC enthusiasts will find their hard-earned cash is likely better spent elsewhere — for example, anything from our best graphics cards rankings.

Moore Threads Technology, a Chinese GPU developer that was blacklisted by the U.S. government this week, is testing its next-generation graphics processors for client PCs and datacenters, according to Löschzwerg from 3D Center over at Devicehunt.com. 

The products in question are the MTT S90 for client computers and the MTT S4000 for datacenters. The devices have 0301 and 0323 Device IDs, respectively, which may indicate that they're the 3rd Generation GPUs from Moore Threads. 

We don't know much about these GPUs, though, given the new Device IDs, we can only speculate that we're dealing with a brand-new microarchitecture that will succeed the company's Chunxiao GPU series. The Chunxiao series includes the MTT S70, MTT S80, and MTT S4000 products, which can barely compete against the best graphics cards from AMD, Intel, and Nvidia.

Device Hunt monitors PCI ID and USB ID repositories to keep tabs on new hardware which identifications emerge in one database or another. Typically, this happens when companies begin to test their new chips and/or their drivers, so it's highly likely that Moore Threads already has silicon of its 3rd Generation GPU on hand. At this point it is impossible to tell how well that graphics processor works and how good its drivers will be. 

But maybe it doesn't matter. As of yesterday, Moore Threads is on the U.S. Entity List, which limits its access to American technologies. Since virtually all chips made at TSMC and other foundries use technologies or tools designed in the U.S., chipmakers have to apply for an export license from the U.S. Department of Commerce's Bureau of Industry and Security — which is reviewed with a presumption of denial.

Without access to TSMC or other foundries, such as Intel Foundry Services or Samsung Foundry, Moore Threads will have to try to persuade SMIC to build its GPUs. It's unclear whether the latter would be able to produce enough GPUs using its 7nm-class process technology, considering it also serves Huawei's Hisilicon Kirin 9000S and possibly other mobile SoCs — but it doesn't look like Moore Threads will have much of a choice.

"As described in an upcoming amendment to regulations regarding advanced computing items and supercomputer and semiconductor end use, advanced computing ICs can be used to provide artificial intelligence capabilities to further development of weapons of mass destruction, advanced weapons systems, and high-tech surveillance applications that create national security concerns," an explanation by the U.S. Department of Commerce reads. "This activity is contrary to U.S. national security and foreign policy interests under § 744.11 of the EAR."

Under the terms of the latest export rules imposed by the U.S. government, AMD, Intel, and Nvidia can no longer ship a number of their high-performance processors to China and a number of other countries without an export license from the U.S. Department of Commerce. In fact, the restrictions are so severe that shipments of Nvidia's AD102 processors are also restricted, which may have an impact on the supply of GeForce RTX 4090 graphics cards.

Given the demand for Nvidia's AI GPUs across the globe, the company does not expect its financial results in the near term to be affected by the new export rules. Meanwhile, it remains to be seen how these new export rules affect the production and prices of GeForce RTX 4090-based graphics cards, which are generally made in China. In a bid to comply with the new export rules, Nvidia will have to initiate production of GeForce RTX 4090 and other AD102-based products outside of China.

Transferring production to Taiwan is perhaps not a problem for the vast majority of brands as the most of them are headquartered in Taiwan. There is one big exception, though. Colorful, which is one of Nvidia's major customers and which happens to be one of the world's largest graphics card manufacturers, only operates in China. Nvidia may attempt to supply as many AD102 GPUs to its partner as possible in the coming weeks, though it remains to be seen how this affects the supply of GeForce RTX 4090 products in other countries. 

Starting from November 16, 2023, Nvidia will be unable to ship its A100, A800, H100, H800, L40, L40S, and GeForce RTX 4090 cards and modules for AI and HPC computing to China, Saudi Arabia, the United Arab Emirates, and Vietnam without an export license from the U.S. Department of Commerce's Bureau of Industry and Security. 

All the aforementioned products, except the GeForce RTX 4090, are data center GPUs for AI, HPC, and cloud applications. The GeForce RTX 4090 is the best graphics card money can buy, assuming price isn't a limiting factor, but since the new restrictions curb exports of high-performance processors in general, it falls under the new regulations. 

"The licensing requirement may impact the Company's ability to complete development of products in a timely manner, support existing customers of covered products, or supply customers of covered products outside the impacted regions, and may require the Company to transition certain operations out of one or more of the identified countries," a statement by Nvidia in its filing with the SEC reads.

In a bid to thwart the development of artificial intelligence (AI) technologies in China, the U.S. government this week imposed further restrictions on exports of AI and HPC GPUs to the People's Republic and blacklisted two Chinese GPU developers, reports Reuters. As a result of the new controls, Nvidia will be unable to sell its A800 and H800 AI and HPC GPUs to Chinese entities, while Biren and Moore Threads will likely lose access to advanced production nodes.

Last year the U.S. banned exports of processors that could enable Chinese entities to build supercomputers with performance of over 100 FP64 PetaFLOPS or over 200 FP32 PetaFLOPS within 41,600 cubic feet (1178 cubic meters). As a result, Nvidia lost the ability to sell its A100, A100X, and H100-series products to China-based companies and had to build the A800 and H800 GPUs with cut-down communication capabilities specifically for China to comply with restrictions. Intel did something similar with its Gaudi 2 solutions for China. But while the restrictions limited supercomputers and large deployments, they never limited the performance of processors sold to Chinese companies. This is being addressed by the new rules.

The new regulations introduced on Tuesday remove restrictions on communication speed and prioritize computing performance instead. The U.S. government now restricts a specified level of performance density within a specific silicon volume, according to Reuters. This adjustment is expected to stop the sales of Nvidia’s A800 and H800 chips in the China market. They are also meant to restrict performance of multi-chiplet solutions shipped to Chinese companies, which further restricts sales of things like AMD's Instinct MI300.

"These export controls are intended to protect technologies that have clear national security or human rights implications," said Gina Raimondo, the U.S. Commerce Secretary, on a call with reporters, reports Associated Press. "The vast majority of semiconductors will remain unrestricted. But when we identify national security or human rights threats, we will act decisively and in concert with our allies."

Apparently, it was not enough for the U.S. government to restrict sales of American AI and HPC processor to Chinese companies. The Biden administration also put Biren Technology and Moore Threads, two GPU developers from China, onto its Entity List. Details about the move are scarce, but typically inclusion on an Entity List restricts a company's access to American technology. As a result, any company — whether American or foreign — that supplies American-developed technology or uses American technology to provide services or supply goods to such entities has to obtain an export license from the U.S. Department of Commerce's Bureau of Industry and Security. Such licenses are reviewed with presumption of denial and are rarely granted.

"The goal is the same goal that has always been, which is to limit PRC access to advanced semiconductors that could fuel breakthroughs in artificial intelligence," said Raimondo, reports Axios. "It is true that AI has the potential for huge societal benefits. But it also can do tremendous and profound harm if in the wrong hands and in the wrong militaries."

In the cases of Biren Technology and Moore Threads, it is likely that both will lose access to TSMC's advanced process technologies, which will essentially stop their businesses until they manage to produce their GPUs somewhere else. For example, at China-based SMIC. Such a transition will likely take quite some time to be completed.

With U.S. restrictions on high-end artificial intelligence (AI) and high-performance computing (HPC) exports to China, Chinese GPU developers see a golden opportunity to fill the void in their domestic market. For now, the market of AI hardware is dominated by Nvidia, but as Chinese GPU vendors polish their hardware and software, they may carve out more than a niche, reports DigiTimes.

Chinese firms such as ILuvatar CoreX, Moore Threads, and BirenTech are actively collaborating with local cloud computing providers like Bidu to run their LLM services. Other companies like DenglinAI, Vast AI Tech, and MetaX have developed products that run large language models from the start. For companies like Moore Threads, this move indicates a shift in focus from gaming hardware to a more profitable data center business. 

Interestingly, some Chinese enterprises are exploring alternative routes to compete in the AI-accelerated computing sector. For instance, Alibaba subsidiary T-Head, and Enflame, are investigating in non-GPU solutions as a way to gain a foothold in the market. This diversification strategy could offer them a unique selling proposition and make them more competitive. Meanwhile, since Alibaba develops chips for its own data centers, they are poised to be a success if they deliver the right performance at the right power.

The current situation is a mixed bag for these Chinese companies. While Nvidia's market dominance serves as an educational platform that raises awareness about the importance of compute GPUs for AI and HPC, it also leaves little room for other players, especially in the global arena. However, the firms are optimistic, as they believe Nvidia's success is essentially laying the groundwork for them as the market shifts from CPUs to compute GPUs.

In fact, the most recent quarter results of Intel and Nvidia prove that, for now at least, GPUs can better address the emerging AI sector. Nvidia's revenue totaled $13.5 billion in Q2 FY2024, whereas Intel's revenue totaled $12.9 billion. Those numbers might seen pretty close, but Nvidia's data center revenue reached $10 billion, which is more than double that of Intel's data center revenue of $4 billion.

Japanese tech site PC Watch has managed to get its hands on a Moore Threads MTT S80 graphics card. This card uses a GPU that, despite some obfuscation deployed by Moore Threads management, appears to use the Imagination Technologies PowerVR architecture. This isn't going to make the list of the best graphics cards, not even if we roll back the clock to 2015, but it's good to start to see cards like this tested outside of China in benchmarks and games we're familiar with.

Though we're pretty sure that the PowerVR architecture is behind the card, PC Watch reported from Moore Threads verbatim regarding the specs of the MTT S80. So, the card uses the Chunxaio architecture, employing 4096 MUSA cores. Other specs include the GPU's clock speed of 1.8 GHz, and its peak performance of 14.2 TFLOPS. There's a generous 16GB of GDDR6 14 Gbps memory onboard the sample tested by PC Watch, and that connects to the GPU via a 256-bit bus for 448 GB/s bandwidth.

The MTT S80 is relatively power hungry, with a TGP (total graphics power) of 255W. That's probably why it includes a triple fan design. Also interesting is that the card uses a PCI Express Gen5 x16 connector. These raw specs don't tell the full story, of course, and driver support could still be a major factor in performance. But let's see what PC Watch found in its testing.

Before we look at the benchmarks and gaming tests, please note that PC Watch found there were lots of current games that wouldn't run on the MTT S80 - even using a supported motherboard, OS and CPU. DX12 and Vulkan games were insurmountable hurdles for this card, but some DX11 titles could run with varying degrees of success. Modern benchmarks faced a similar issue, with the most current version of 3DMark stable and usable being 3DMark 06.

Above you can see that the MTT S80 fares very badly when put up against even modest competition like Nvidia's GeForce GTX 1050 Ti, a budget GPU that debuted in 2016. On paper, the GTX 1050 Ti is woefully outmatched — it offers up 2.1 teraflops of compute, has 4GB of GDDR5 memory, and 112 GB/s of bandwidth with a 75W TGP. Even with such weak competition, the Moore Threads graphics card falls flat.

The MTT S80 fared best in DX9 graphics comparisons. It was still significantly behind the GTX 1050 Ti but not completely trounced. Or... well, it was completely trounced, but the average performance lead in the DX9 tests was 'only' 86% — so not quite double the performance.

Things were much worse for the MTT S80 when PC Watch looked at a selection of DX11 games. Besides some rendering errors, performance was very poor and some games effectively failed to work at all with single digit fps. Even if we discount Asetto Corsa where the Nvidia GPU was 90 times faster, the average lead in the DX11 games was still 188%, nearly triple the performance. One percent lows were also frightful on the MTT S80.

Poor performance and compatibility isn't the end of the sorry tale of the Moore Threads graphics card, unfortunately. As it stands, the card sucks up a lot of watts for very little. The MTT S80 on average consumed 142W, while the GTX 1050 Ti averaged just 60W.

In terms of performance per watt, even discounting Asetto Corsa (again), the MTT S80 managed just 0.33 fps/W while the GTX 1050 Ti averaged 1.86 fps/W. That makes the old Pascal GPU over five times as efficient.

There's a clear disconnect between the raw specs of the MTT S80 and its real-world results. On paper, the MTT S80 has four times as much memory, four times the memory bandwidth, and nearly seven times the raw FP32 compute. It's nowhere near reaching that theoretical level of performance.

PC Watch seems to think that the Moore Threads graphics card's major issue is with drivers, so it has some hope that things will continue to improve over the coming months. For now, the MTT S80 is not for gamers, curious developers, or graphics card collectors.

Back in February we reported on Korean TechTuber BullsLab Jay's video featuring the same MTT S80 graphics card. At that time, gaming tests were restricted to DX9 titles due to platform / driver immaturity. The fact that the MTT S80 can now at least try to run some DX11 games shows progress with the drivers, but there's still a long way to go. DirectX 12 and Vulkan games are also not currently supported.

Moore Threads has introduced its new midrange graphics card, the MTT S70, which is powered by a cut-down version of its Chunxiao graphics processor and is equipped with 7GB of GDDR6 memory. The product will offer FP32 compute throughput of a little bit lower than that of the GeForce RTX 3060 in boost mode, though its real-world performance is something that remains to be seen.

The Moore Threads MTT S70 is based on the Chunxiao GPU with 3584 stream processors that operates at 1.60 GHz (down from 4096 SPs at 1.80 GHz in case of the MTT S80) and offers FP32 compute performance of 11.2 TFLOPS, which is slightly below peak performance of Nvidia's GeForce RTX 3060 (12.74 TFLOPS). The board comes with a large cooling system and has three DisplayPort 1.4a and one HDMI 2.1 output, which is in line with the flagship MTT S80.

The thing that catches the eye about the MTT S70, though, is its rather odd memory configuration: the board carries 7GB of memory with 392 GB/s bandwidth, which suggests that its memory interface was cut down to 224-bits. Apparently, Moore Threads also decided to reduce memory capacity rather significantly, so the board comes with 7GB of memory, which is not enough for many games by today's standards.

While compute performance and memory capacity are extremely important for overall performance of a GPU, things like overall architecture efficiency and drivers have a great influence on performance of a graphics card. Moore Threads's current flagship MTT S80 has FP32 throughput comparable to that of GeForce RTX 3060 Ti, but it not only failed to match GeForce RTX 3060 in games that its drivers support, but it also failed to outperform Nvidia's GeForce GT 1030 due to poor drivers. 

In any case, at this point even the range-topping and rather powerful MTT S80 probably has no chances to join the ranks of the best graphics cards and so the MTT S70 is also unlikely to do this. Hopefully, Moore Threads fixes its drivers rather sooner than later. During the event the company demonstrated a slide detailing MTT S80's performance improvements from Q4 2022 to Q2 2023.

Just like other GPU designers, Moore Threads, a rising star of the Chinese semiconductor industry, uses cut-down versions of its flagship GPUs to address lower-end market segments. The MTT S70 received a rather significant cutdown, which will enable the GPU developer to ship a boatload of these graphics cards if there is enough demand.

Now that AMD, Intel, and Nvidia cannot sell their top-of-the-range compute GPUs to customers in China without permission of the U.S. government, and Chinese developers of GPUs for artificial intelligence (AI) and high-performance computing (HPC) applications do not have access to leading-edge production capabilities, smaller GPU designers have a chance to capitalize on various types of generative AI, reports DigiTimes. 

The rise of the Chinese chip design sector in the recent years has formed two vectors of domestic GPU development: datacenter GPUs designed to address AI and HPC megatrends and classic GPUs designed primarily for client PCs yet capable of addressing some datacenter-specific workloads. It looks like the latter have better chances for success in the current situation.

Developers of Small GPUs Can Address Big Things

Innosilicon, Jingjia Microelectronics, and Moore Threads are perhaps the most well-known Chinese developers of gaming graphics processors. Gaming graphics depends on single-precision floating point (FP32) compute throughput, so GPUs from Innosilicon, Jingjia, and Moore's Threads support this data format. For now, these GPUs can hardly claim a place among the best graphics cards. Furthermore, to address various artificial intelligence and machine learning applications, these companies are going to need to tweak their hardware to support lower-precision data formats (think FP16, BF16/8, INT8, INT 4, etc.) as well as specific instructions for matrix and vector processing. Some have already done this, while others have yet to. 

Jingjia has been offering gaming GPUs based on its own architectures since 2014 and its latest JM9-series GPUs from 2021 promise to offer performance levels similar to Nvidia's GeForce GTX 1080. For now, Jingjia's GPUs cannot address AI/DL/ML applications, but the company told DigiTimes that it was working on AI-capable GPUs for a variety of applications, such as speech recognition and natural language processing, but did not elaborate. 

Innosilicon introduced its 1st Generation Fenghua (Fantasy) discrete graphics processor featuring a PowerVR microarchitecture from Imagination Technologies in late 2021, followed up with 2nd Generation Fantasy GPU for low-power applications in mid-2022, and announced development of its 3rd Generation Fantasy with ray tracing support last year. Innosilicon's original Fantasy GPU supports both FP32 and INT8, while its drivers support modern application programming interfaces for compute, including DirectX, Vulkan, OpenCL, Caffe 1.0, TensorFlow 1.1.2, and ONNX. 

Hardware from Moore Threads is perhaps better suited for AI. The company's latest Chunxiao graphics processor supports FP32, FP16, and INT8 precision and, assuming that it also supports appropriate instruction sets, can address at least some AI/DL/ML workloads. The company also says that its MTVerse platform can enable developers to build applications for big data, AI training and reasoning, speech recognition, and visual recognition, among other things.

Some Background

Total available market of graphics processing units in China reached $4.739 billion, making up 18.7% of the global market share, according to data from VMR cited by DigiTimes. It is unclear whether datacenter GPUs are included in $4.739 billion. The market is projected to increase to US$34.56 billion by 2027, with a compound annual growth rate (CAGR) of 32.8% over the seven-year period. 

There are about 10 developers of GPUs in China. Two of them — Biren Technology and Tianshu Zhixin Semiconductor — are focused purely on processors for AI and HPC applications and their GPUs are not exactly meant to process graphics. But these processors need leading-edge fabrication technology to be made and TSMC needs an export license from the U.S. government to produce GPUs for Biren and Tianshu Zhixin. 

Meanwhile, there are plenty of developers of more or less universal GPUs — in addition to Jingjia, Innosilicon, and Moore Threads — for rendering games, but which can also address AI and technical computing applications if they gain appropriate hardware capabilities.  

Up until recently Chinese designers of gaming GPUs were not inclined to build AI and HPC-oriented processors that would rival solutions from big companies like Nvidia, Intel, AMD, Biren, and Tianshu Zhixin. Now that the future of China-based AI and HPC GPU developers is uncertain and the abilities of AMD, Intel, and Nvidia to address Chinese customers are limited, they may review their plans and come up with chips that can address China's needs for AI/DL/ML hardware.

China's Moore Treads MTT S80 graphics card has been tested in a suite of games by TechTuber BullsLab Jay. They say these GPUs only come up for sale occasionally in small batches, so they were fortunate to be able to grab one.

In brief, the 12nm, 4096 core GPU with 16GB of GDDR6 and a TGP of 250W is disappointing in gaming, as it was consistently outgunned by the GT 1030 2GB with a 30W TGP. Moreover, the card seemed to be restricted to DX9 gaming at the time of review (it is advertised as DX11-capable) and its AV1 codec acceleration was also lacking. Last but not least, BullsLab Jay’s investigation shows that the MTT S80 relies on a PowerVR-based GPU architecture – so its mysterious Chunxaio GPU with MUSA cores has now been unmasked.

In the BullsLab Jay video, you can see the Moore Threads graphics card unboxed, tested and disassembled. If you open the link you can watch it with closed-captioned in English, but we have embedded the benchmarks-only video from the TechTuber’s ENG channel below. So you can check out some of the benchmarking action without any fiddling with caption controls.

The 250W MTT S80 has a bulky triple fan cooler, but once removed you can see the PCB is only about two fans in length. Contrast this design with the 30W GT 1030 which can run with a passive heatsink. The other card in the tests, a 120W GTX 1060, offered stratospheric gaming performance compared to the MTT S80, but it's also usually manufactured in single- or two-fan versions.

Getting to the benchmarks, the titles under comparative tests were limited to DX9 games, as that is all that worked on the Moore Threads review card / driver. The architecture is claimed to support DX11, so it looks like some driver updates will be required. In the video's summary charts, the MTT S80 is often half as fast as the aging low-power GT 1030. Only an outlier or two indicates the MTT S80 has some potential.

Speaking of unfulfilled potential, the touted AV1 video codec support claims are also ringing hollow right now. Several tests showed AV1 decoding was tackled by the CPU, however the card was seen to shove VP9 Codec video processing to the GPU in YouTube testing.

BullsLab Jay dug through old Moore Threads China presentations and concluded that the MTT S80 uses Imagination Technologies' PowerVR architecture. Moore Threads have not been very upfront about this, and some had hoped that it was taking a different tack to the Innosilicon Fantasy (Fenghua) line, which also uses PowerVR IP. (And again, it wasn’t very up-front in letting people know about the underlying architecture).

The Moore Threads MTT S80 graphics card sells for the equivalent of approximately $430 in China, bundled with an Asus TUF B660M motherboard. If we take away the cost of the board, the graphics card is roughly $260. Perhaps it will start to show more of its FP32 potential as drivers mature and these PowerVR cards for PCs will become more effective challengers.

Right now, with demonstrably faster GT 1030 selling for $80 or less, the MTT S80 isn't exactly making a good case for itself at more than three times that price. It's possible that performance will improve substantially with driver updates, just as we saw with Intel over the past several months. But of course, that will depend on how much time and effort the company plans to spend on driver improvements, as well as the limitations of the architecture. 

China doesn't have many homebrew graphics cards, so expectations were high when graphics card manufacturer Moore Threads revealed the MTT S80. It's hard to tell whether the MTT S80 has what it takes to compete with the best graphics cards, but hopefully, graphics card collector Löschzwerg's latest benchmarks can provide some insight.

While Moore Threads is green behind the ears, the company has strong leadership. Former Nvidia global VP and China GM Zhang Jianzhong founded Moore Threads in 2020, so the Chinese fabless semiconductor company is a newcomer to the graphics card game. Besides being China's domestic graphics card, the MTT S80 has garnered a fair amount of hype outside the country since it's the first PCIe 5.0 gaming graphics card to hit the market. Nvidia's GeForce RTX 40-series (Ada Lovelace) and AMD's Radeon RX 7000-series (RDNA 3) products are still on PCIe 4.0.

The MTT S80 is the successor to the MTT S60 and still leverages the same MT Unified System Architecture (MUSA) architecture. It supports modern APIs, including CUDA, DirectX, OpenCL, OpenGL, and Vulkan. The Chunxiao GPU, manufactured under the 12nm process node, powers the MTT S80. Like Nvidia and AMD, the MTT S80 embraces AV1 encoding support in addition to other popular formats, such as H.264, H.265, and VP9.

The MTT S80, which supports PCIe 5.0 x16, wields 4,096 MUSA cores operating at 1.8 GHz to offer up to 14.4 TFLOPs of FP32 performance. This places the Chinese graphics card between the GeForce RTX 3060 (12.7 TFLOPs) and GeForce RTX 3060 Ti (16.2 TFLOPs) or, alternatively, the Radeon RX 6750 XT (13.3 TFLOPs) and Radeon RX 6800 (16.2 TFLOPs). Moore Threads outfits the MTT S80 with 16GB of 14 Gbps GDDR6 memory across a 256-bit memory interface. This arrangement is suitable for a maximum theoretical memory bandwidth of up to 448 GBps, on equal footing with the GeForce RTX 3060 Ti.

The MTT S80 features your typical dual-slot design with a triple-fan cooling solution. Of course, it's a gaming graphics card; some RGB eye candy is inevitable. The graphics card utilizes an 8-pin EPS connector, yes, the kind that you find on motherboards. It has a 255W TBP (total board power), and one EPS connector supplies up to 300W. The MTT S80 has the same outputs as Nvidia's flagship GeForce RTX 4090. In addition, you receive three DisplayPort1.4a outputs and one HDMI 2.1 port to support up to four 8K displays.

To say that benchmarking the MTT S80 was difficult is an understatement. Löschzwerg emphasized that the graphics card's performance and driver were wonky. The GPU utilization wasn't optimal, and the graphics card rarely showed its full potential, suggesting a lack of driver optimization. Tessellation doesn't work with the current driver and causes crashes on Unigine Heaven, 3DMark 11, and Final Fantasy XIV: Endwalker. Surprisingly, Resizable BAR works fine.

Moore Threads MTT S80 Benchmarks

The MTT S80 delivered between 16 FPS and 31 FPS (actual gameplay) in Crysis at 1080p on the high preset with DirectX 9. In F.E.A.R., with 1080p and maximum preset, the graphics card ran the game between 54 FPS and 186 FPS. Meanwhile, the average framerate for Half-Life 2: Lost Coast at 1080p and maximum settings was 132.2 FPS.

The MTT S80 finished the Final Fantasy XIV: Endwalker benchmark with a score of 4,312 points. Unfortunately, the reviewer used custom settings so we couldn't compare it to our results. However, the score slides between the 4,000 to 5,999 range, equivalent to standard performance, meaning the graphics card has what it takes to run the game on default settings. The Arc A770 scored 18,674 points in the same benchmark, outpacing the MTT S80 by 100%.

The results revealed that the Arc A770 outperformed the MTT S80 by over 170% in 3DMark06 on both presets. Intel's graphics card also delivered 155% higher performance in 3DMark03. With the more demanding preset, the Arc A770 still beat the MTT S80 by a 54% margin.

The MTT S80 scored 551 points in Unigine Heaven at 1080p 8xAA on the ultra preset. Another Twitter user provided the scores for the GeForce GTX 680 and the Radeon HD 7950 Boost. The former had a 114% higher score, while the latter showed a 101% better score than the MTT S80.

The power metrics are a real shocker, though. The test system, which comprises the Core i5-10400, ASRock B560M-HDV, and 32GB (2x16GB) of DDR4-3200 memory, idles around 22W. With the MTT S80, the idle power was about 131W, conveying that the graphics card consumes 109W while idling. On the other hand, the peak system power consumption was 315W, so the MTT S80 consumed 293W. Again, the values are ridiculously high. For example, even the enthusiast-grade GeForce RTX 4070 Ti is a 285W graphics card.

With the driver's current state, the MTT S80 is potentially leaving some performance on the table. Officially, the graphics card supports around 20 DirectX games, but performance is a hit or miss. The graphics card performs better on DirectX 9 titles than on DirectX 11. However, there's still a long way to go regarding game compatibility.

The MTT S80 retails for 2,999 yuan or $442.65 on JD.com, a popular Chinese online retailer. Unfortunately, it won't be a suitable option since gamers can still purchase Intel, AMD, or Nvidia graphics cards in China. However, if restrictions on exports to China get harsh, the MTT S80 could gain relevance.

Moore Threads, a China-based GPU developer, raised around $215.4 million in Series B funding to finance its continuous multi-functional GPU research and development, reports DigiTimes. The fund-raising signals investors' confidence in the rise of China GPU development in general fueled by the country's need for AI/ML accelerators as well as graphics processors for gaming. 

Since its inception in late 2020, Moore Threads has introduced two graphics processors compatible with the company's MUSA computing platform — Sudi and Chunxiao — and multiple add-in-boards (AIBs) built around that base design. Both multi-functional GPUs are designed for numerous markets, including gaming, cloud computing, artificial intelligence/machine learning, FP32 high-performance computing, and virtual desktop infrastructure (VDI). Pursuing so many market opportunities lets Moore Threads attract investors with various backgrounds.  

In its Series B funding round, Moore Threads secured $215.4 million from an investment fund of China Mobile and Hexe Health Insurance. China Mobile can take advantage of Moore Threads GPU's AI/ML, gaming, and cloud computing capabilities, whereas Hexe Health Insurance is probably interested primarily in the acceleration of its artificial intelligence workloads. 

Meanwhile, the company received seed funding and A Series funding rounds from ByteDance (Tik Tok owner), Tencent, Shenzhen Capital Group (SCGC), Sequoia China, and GGV Capital. The Series A funding round brought the company around $313 million.  

Moore Threads' latest Chunxiao is a rather complex GPU containing 22 billion transistors and packing 4,096 stream processors, 128 tensor cores, 256 texture units, and 256 render outputs clocked at 1.80 GHz. The unit supports FP32, FP16, and INT8 precision for various workloads and comes with a video engine that supports AV1, H.264, and H.265 codecs for up to 8K videos as well as can decode up to 32 streams at a 1080p30 resolution. 

Meanwhile, the Chunxiao-based MTT S80 16GB graphics card yet has to demonstrate its full potential in video games popular in Europe and the U.S. as it currently fails to beat even Nvidia's GeForce RTX 3060 despite its higher complexity and compute performance (14.4 FP32 TFLOPS vs 12.7 FP32 TFLOPS) due to imperfect drivers. 

Yet, the MTT S80 as well as its datacenter-oriented MTT S3000 sibling can be used to play games popular in China as well as for remote Android game rendering, VDI, various artificial intelligence, machine learning, light high-performance computing (HPC) that only requires FP32 precision, and video streaming workloads. Essentially, Moore Threads' Chunxiao GPU meets all the primary goals of a Chinese GPU developer — address China's needs in AI/ML, gaming, cloud computing, and video streaming. 

Moore Threads has proven that it can build rather complex and capable graphics processors, but now the company's main challenge is to ensure that these GPUs can address all the workloads they are intended for, which means developing drivers and the rest of the software stack. Market leader Nvidia has spent years polishing off its drivers to address markets that Moore Threads wants to address. Will Moore Threads be able to do the same in a shorter amount of time is something only time will tell. But at least investors seem to be confident that over time the company will be able to at least partially replace Nvidia in China.

The number of GPU startups in China is extraordinary as the country tries to gain AI prowess as well as semiconductor sovereignty, according to a new report from Jon Peddie Research. In addition, the number of GPU makers grew worldwide in recent years as demand for artificial intelligence (AI), high-performance computing (HPC), and graphics processing increased at a rather unprecedented rate. When it comes to discrete graphics for PCs, AMD and Nvidia maintain lead, whereas Intel is trying to catch up. 

18 GPU Developers

Tens of companies developed graphics cards and discrete graphics processors in the 1980s and the 1990s, but cut-throat competition for the highest performance in 3D games drove the vast majority of them out of business. By 2010, only AMD and Nvidia could offer competitive standalone GPUs for gaming and compute, whereas others focused either on integrated GPUs or GPU IP.

The mid-2010s found the number of China-based PC GPU developers increasing rapidly, fueled by the country's push for tech self-sufficiency as well as the advent of AI and HPC as high-tech megatrends. 

In total, there are 18 companies developing and producing GPUs, according to Jon Peddie Research. There are two companies that develop SoC-bound GPUs primarily with smartphones and notebooks in mind, there are six GPU IP providers, and there are 11 GPU developers focused on GPUs for PCs and datacenters, including AMD, Intel, and Nvidia, which design graphics cards that end up in our list of the best graphics cards.

In fact, if we added other China-based companies like Biren Technology and Tianshu Zhixin to the list, there would be even more GPU designers. However, Biren and Tianshu Zhixin are solely focused on AI and HPC for now, so JPR does not consider them GPU developers. 

China Wants GPUs

Being the world's second largest economy, China inevitably competes against the U.S. and other well-developed countries in terms of pretty much everything, including technology. China did a lot to lure engineers from around the world and make it worthwhile to establish various chip design startups in the country. In fact, hundreds of new IC design houses emerge in China every year. They develop all kinds of things from tiny sensors to complicated communication chips, thus enabling the country's self-sufficiency from Western suppliers.  

But to really jump on the AI and HPC bandwagon, China needs CPUs, GPUs, and special-purpose accelerators. When it comes to computing, it is impossible for Chinese companies to leave behind long-time CPU and GPU market leaders any time soon. Yet, it is arguably easier and perhaps more fruitful to develop and produce a decent GPU than try to build a competitive CPU.  

"AI training was the big motivator [for Chinese GPU companies], and avoidance of Nvidia's high prices, and (maybe mostly) China's desire for self-sufficiency," said Jon Peddie, the head of JPR. 

GPUs are inherently parallel, which means there are loads of compute units inside that can be used for redundancy, which makes it easier to get a GPU up and running (assuming that per transistor costs are relatively low and overall yields are decent). Also, since GPUs are fundamentally parallel, it is easier to parallelize them in scale-out manner. Keeping in mind that China-based SMIC does not have production nodes as advanced as those of TSMC, this way of performance scaling looks good enough. In fact, even if Chinese GPU developers lose access to TSMC's advanced nodes (N7 and below), at least some of them could still produce simpler GPU designs at SMIC and address AI/HPC and/or gaming/entertainment market. 

From China's perspective as a country, AI and HPC-capable GPUs may be arguably more important than CPUs too since AI and HPC can enable all-new applications, such as autonomous vehicles and smart cities as well as advanced conventional arms. The U.S. government of course restricts exports of supercomputer-bound CPUs and GPUs to China in a bid to slowdown or even constrain development of advanced weapons of mass destructions, but a fairly sophisticated AI-capable GPU can enable an autonomous killer drone, and drone swarms represent a formidable force, for instance. 

GPU Microarchitecture Is Relatively Easy, Hardware Design Is Expensive

Meanwhile, it should be noted that while there are a bunch of GPU developers, only two can actually build competitive discrete GPUs for PCs. That is perhaps because it is relatively easy to develop a GPU architecture, but it is truly hard to properly implement it and to design proper drivers. 

CPU and GPU microarchitectures are essentially at the intersection of science and art. They are sets of sophisticated algorithms that can be developed by rather small groups of engineers, but they might take years to develop, says Peddie.  

"[Microarchitectures] get done on napkins and white boards," said Peddie. "[As for costs] if it is just the architects themselves, that [team] can be as low as one person to maybe three – four. [But] architecture of any type, buildings, rocket ships, networks or processors is a complicated chess game. Trying to anticipate where the manufacturing process and standards will be five years away, where the cost-performance tradeoffs are, what features to add and what to drop or ignore is very tricky and time-consuming work. […] The architects spend a lot of time in their head running what-if scenarios — what if we made the cache 25% bigger, what if we had 6,000 FPUs, should we do a PCIe 5.0 I/O will it be out in time." 

Since microarchitectures can take years to develop and they require talented designers, in a world where time-to-market is everything, many companies license an off-the-shelf microarchitecture or even a silicon-proven GPU IP from companies like Arm or Imagination Technologies. For example, Innosilicon — a contract developer of chips and physical IP — licenses GPU microarchitecture IP from Imagination for its Fantasy GPUs. There is another China-based GPU developer, which uses a PowerVR architecture from Imagination. Meanwhile, Zhaoxin uses a highly reiterated GPU microarchitecture it acquired from Via Technologies, which inherited it from S3 Graphics. 

The cost of developing a microarchitecture may vary, but it is relatively low when compared to the costs of a physical implementation of a modern high-end GPUs. 

For years, Apple and Intel, both companies with plenty of engineering talent, relied on Img for their GPU designs (Apple still does to a certain extent). MediaTek and other smaller SoC suppliers rely on Arm.  Qualcomm used ATI/AMD for an extended period, and Samsung uses AMD after several years of trying to design their own graphics engine.

Two of the new Chinese companies have hired ex AMD and Nvidia architects to start their GPU companies, and another two use Img. Time to market and learning the skills of being an architect, what to worry about, and how to find a fix is a very time consuming process.  

"If you can go to a company that already has a design and have been designing for a long time, you can save a boatload of time and money – and time to market is everything," said the head of Jon Peddie Research. "There are just so many gotchas. Not every GPU designed by AMD or Nvidia has been a winner. [But] a good design lasts a couple of generations with tweaks." 

Hardware implementation and software development are prohibitively expensive with new production nodes. International Business Times estimates that the design costs for a fairly complex device made using 5nm-class technology exceeds $540 million. These costs will triple at 3nm. 

"If you include layout and floor plan, simulation, verification, and drivers then the [GPU developer] costs and time skyrocket," explained Peddie. " The hardware design and layout is pretty straight forward: get one trace wrong and you can spend months tracking it down." 

There are just a few companies in the world that can develop a chip featuring the complexity of modern gaming or compute GPUs from AMD and Nvidia (46 billion — 80 billion transistors), yet China-based Biren could do something similar with its BR104 and BR100 devices (we speculate that the BR104 packs some 38.5 billion transistors).  

Thoughts 

Despite prohibitive costs, eight out of the 11 PC/datacenter GPU designers are from China, which speaks for itself. Perhaps we won't see a competitive discrete gaming GPU from anyone except huge American companies in the near future. That's partly because its hard and time consuming to develop a GPU, and to a large degree it requires a prohibitively expensive hardware implementation for these high-complexity GPUs. Whether or not China can field competitive entrants remains to be seen, but any failure won't stem from a lack of trying. 

A review posted at EXPreview gives us our first look at Chinese GPU maker Moore Thread's new mid-range gaming competitor, the MTT S80 graphics card. This GPU is one of Moore Thread's most powerful graphics cards to date, packing a triple fan cooler setup and theoretically competing with Nvidia's RTX 3060 and RTX 3060 Ti, some of the best graphics cards available, even close to two years after they first launched.

For the uninitiated, Moore Thread is a Chinese GPU manufacturer that was established just two years ago, in 2020. The company has reportedly tapped some of the most experienced minds in the GPU industry, hiring experts and engineers from Nvidia, Microsoft, Intel, Arm and others. Moore Thread's aim is to produce domestic (for China) GPU solutions completely independent from Western nations. These are supposed to be capable of 3D graphics, AI training, inference computing, and high-performance parallel computing capabilities, and will be used in China's consumer and government sectors.

The MTT S80 graphics card was developed with Moore Thread's 'Chunxaio' GPU architecture, which supports FP32, FP16, and INT8 (integer) precision compute, and is compatible with the company's MUSA computing platform. The architecture also employs a full video engine with H.264, H.265 (HEVC), and AV1 codec support, capable of handling video encoding and decoding at up to 8K.

The MTT S80 comes with a fully unlocked Chunxaio GPU core, featuring 4096 MUSA cores, and 128 tensor cores, and clocked at 1800 MHz. The memory subsystem uses 14 Gbps GDDR6 modules operating on a 256-bit wide bus, with a 16GB capacity. As far as specs go, it at least looks decent on paper.

The GPU operates with a target board power rating of 255W, powered by both a PCIe 5.0 x16 slot, and a single 8-pin EPS12V power connector — yes, it's using a CPU EPS12V rather than an 8-pin PEG (PCI Express Graphics) connector. That's because the EPS12V can deliver up to 300W, and for users that lack an extra EPS12V connector the card includes a dual 8-pin PEG to single 8-pin EPS12V adapter. And for the record, that's more power than even the RTX 3070 requires.

Display outputs consist of three DisplayPort 1.4a connectors and a single HDMI 2.1 port, the same as what you'll find on most Nvidia GeForce GPUs from the RTX 40- and 30-series families. The graphics card has a silver colored shroud, accented by matte black designs surrounding the right and left fans. The cooler features a triple-fan cooler design, with two larger outer fans and a smaller central fan in the middle. The card dimensions are 286mm long and two slots wide.

The RTX 3060 Beats the MTT S80 in Early Benchmarks

Earlier reports on the Chunxaio GPU suggest it can achieve FP32 performance similar to that of an RTX 3060 Ti. The 3060 Ti has theoretical throughput of 16.2 teraflops, while the Chunxaio has a theoretical 14.7 TFLOPS. That's a bit lower than Nvidia, but it does have twice the VRAM capacity, so the hardware at least appears capable of competing with Nvidia's RTX 3060 Ti.

Unfortunately, according to EXPreview, the MTT S80 suffers from very poor driver optimizations. While the MTT S80 is aimed at the RTX 3060 Ti in terms of raw compute performance, in gaming benchmarks Nvidia's RTX 3060 reportedly outpaces the MTT S80 by a significant margin. The problem with the testing is that no comparative charts are provided for the games with RTX 3060 or other GPUs; only text descriptions of the actual gaming performance are provided, while the charts are for an odd mix of titles.

EXPreview ran its benchmarks on an Intel test rig with a Core i7-12700K, Asus TUF B660M motherboard, RTX 3060 12GB Strix, 16GB of DDR4 memory, and an 850W PSU running Windows 10 21H2.

The best potential example of actual gaming performance was Unigine Valley, where the RTX 3060 12GB was anywhere between 2x and a whopping 7.6x faster than the MTT S80 in the DX9 and DX11 tests at 1080p and 4K resolutions. The MTT S60 averaged 26.1 FPS in the 4K DX9 test, while the RTX 3060 spat out a whopping 197.9 FPS. Note also that Unigine Valley was developed and first released in 2009.

3DMark06 — yes, another rather ancient application — showed similar results with the RTX 3060 being 2.5x faster than the MTT S80 on average, at 1080p and 4K resolutions.

The MTT S80 managed to come out on top in some of the synthetic tests, like PCIe bandwidth and pure fill rates. That's understandable on the PCIe link, as the S80's PCIe 5.0 x16 configuration should vastly outpace the RTX 3060's PCIe 4.0 x16 spec. In the OCL Bandwidth Test, the S80 averaged 28.7 read and 42.8GB/s write speeds. The RTGX 3060 managed 18.3GB/s and 14.2GB/s respectively. In 3DMark06's texturing tests, the MTT S80 hit 134.8 GTexels per second in the Single-Texturing Fill Rate test, and 168.5 in the Multi-Texturing test. The RTX 3060 was much slower in the Single-Texturing test, with 59.9 GTexels/s, but came back with 177.3 GTexels/s in the Multi-Texturing test.

Besides these the synthetic and graphics tests, EXPreview did run some actual games: League of Legends, Cross Fire, QQ Speed, QQ Dance, Fantasy Westward Journey, The Great Heroes, Audition, Running Kart, Diablo III, Ultimate Street Fighter IV, Siege, My World, and Need for Speed: Hot Pursuit III. We'll forgive you if you don't recognize some of those, as we don't either. Some are Chinese variants of popular games like Counter-Strike and Final Fantasy.

Performance in a vacuum of course doesn't tell us much. 149 fps at 1080p in League of Legends with max settings, and 128 fps at 4K? Great! How did the RTX 3060 perform? We don't know. Other performance results are equally bad, like hitting the 40 FPS framerate limit in QQ Speed, a rather simplistic looking game.

But the review does mention driver problems, texture corruption, and other issues. It concludes with, "Compatibility needs to be improved and the future can be expected," according to Google Translate. That's hardly surprising for a new GPU company. Hopefully future 'reviews' of the MTT S80 will actually show more real-world comparisons with modern games rather than old tests that have little meaning in today's market.

China-based Moore Treads on Friday began sales of its latest MTT S80 graphics card based on the Chunxiao architecture for client PCs. The add-in-board promises compute performance comparable to Nvidia's GeForce RTX 3060 Ti, but its performance in real games remains unknown. Meanwhile, it is available at JD.com, a store that delivers worldwide. However, there is a catch.

A GeForce RTX 3060 Ti-Like GPU from China

Moore Threads' MTT S80 graphics board is based on the company's Chunxiao graphics processor featuring 4,096 stream processors, 128 tensor cores, 256 texture units and 256 render outputs clocked at 1.80 GHz. In addition, the card carries 16GB of GDDR6 memory with a 14 GT/s data transfer rate connected to the GPU using a 256-bit interface (up to 448 GB/s peak bandwidth). 

From a compute performance point of view, the MTT S80 delivers up to 14.4 FP32 TFLOPS, slightly below the peak compute performance of Nvidia's GeForce RTX 3060 Ti (16.2 FP32 TFLOPS), which is one of the best graphics cards. Meanwhile, the actual performance of the MTT S80 AIB in real games needs to be tested. 

The GPU comes with drivers that support Microsoft's DirectX application programming interface, Khronos Group's OpenGL/OpenGL ES, proprietary MUSA, and multiple specialized APIs. However, Moore Threads openly says that its DirectX support is in its infancy, and only 20 popular titles are initially supported (e.g., Call of Duty, Crossfire, Counter-Strike, Diablo 3, League of Legends, etc.), so do not expect much from the card for now. Meanwhile, the company says that it is working closely with developers of Unity and Unreal Engine as well as designers of actual games to ensure compatibility and performance with the two popular engines. 

The Chunxiao graphics chip also has a capable video engine that can handle AV1, H.264, and H.265 codecs for up to 8K videos. Unfortunately, Moore Threads does not disclose which players support its hardware-assisted video decoding. Meanwhile, the MTT S80 AIB has three DisplayPort 1.4 and one HDMI 2.1 interface. 

Interestingly, the MTT S80 is the world's first client graphics card with a PCIe 5.0 x16 interface, as Nvidia's Ada Lovelace GPUs support a PCIe Gen4 bus. AMD has yet to disclose the host interface of its latest RDNA 3 GPUs in general and Radeon RX 7900 XT and Radeon 7900 XTX in particular.

Only Available in Bundles

The Moore Threads MTT S80 graphics card is now available at JD.com (via VideoCardz) for ￥2999 ($373 without VAT). However, the AIB does not come alone, as it is bundled with the Asus TUF Gaming B660M-PLUS D4 motherboard priced at ￥1029. If you subtract the value of the motherboard, then the graphics board will cost $245 without VAT, which is relatively cheap given its compute capabilities. Still, since the performance of the product is unclear, we can only wonder whether it is actually worth $245. 

Another thing to point out is that according to VideoCardz, the initial release of the MTT S80 will be very limited, and it is unclear how many units were actually made. It is also unclear when the second batch is set to come and how big it is.

A New Kid on the Block?

There are about a dozen known developers of graphics processors in China. Some target only datacenters (including AI and HPC), while others design client GPUs yet only aim local market. Founded in October 2020, Moore Threads designs GPUs for many applications (except HPC) and seems intent on selling its products outside of China. This is why the company makes its MTT S80 available at JD.com, which ships worldwide and works with developers of popular engines and games to ensure compatibility with western titles. 

Will it succeed? That depends on multiple factors, including engineering talent availability, willingness to invest, and ability to execute. Moore Threads is a very young company, and while so far it has released four AIBs (MTT S10, MTT S50, MTT S60, and MTT S2000) with two incoming (MTT S80 and MTT S3000), we can only wonder whether it has enough people to address all of its declared market opportunities. After all, AMD, Intel, and Nvidia employ thousands of people to work on gaming and datacenter GPUs. 

Furthermore, it remains to be seen whether western companies will work closely with a Chinese entity even on such things as video games. Given the strict curbs the U.S. government imposes against Chinese semiconductor and supercomputing sectors, some companies may consider it risky to work with China-based high-tech companies. In fact, there are things that they might deem potentially toxic.

The Chunxiao GPU in its MTT S80 form is aimed at consumer applications and therefore is not designed for any artificial intelligence or high-performance computing applications. Also, Moore Threads has the MTT S3000 32GB server grade board that offers a 15.2 FP32 TFLOPS throughput and supports datacenter-oriented features, such as up to 32-way GPU partitioning as well as SR-IOV PCIe virtualization. Meanwhile, the Chunxiao graphics processor supports FP32, FP16, and INT8 precision. Therefore, assuming that it also supports an appropriate instruction set, it can theoretically be used for AI or even some light HPC workloads that do not need FP64.  

Based on what we know about the Chunxiao graphics processor, we believe it is aimed primarily at gaming and entertainment applications. The datacenter MTT S3000 will likely mainly be used for remote Android games rendering, virtual desktop infrastructure (VDI), and video streaming services.  

Whether the MTT S3000 can actually be used for non-consumer AI and technical computing applications in datacenters efficiently is something we do not know. We have reasonable doubts given the form factor, power consumption, and performance. Yet, if the U.S. further tightens its sanctions against the Chinese high-tech industry, there might be ramifications even for a consumer-oriented company like Moore Threads. 

Moore Threads Intelligent Technology, a major graphics processors developer from China, on Thursday announced its next generation GPU that can be used for gaming, artificial intelligence, and datacenter workloads. The MTT S80 gaming graphics cards as well as MTT S3000 server board promise to deliver compute performance akin to that of Nvidia's GeForce RTX 3060 Ti, though their actual performance in games and professional applications is something that remains to be tested. 

The Chunxiao GPU

Moore Threads's new graphics processor is based around the company's Chunxiao architecture that supports FP32, FP16, and INT8 precision and is compatible with the company's MUSA computing platform and application programming interface (it can work with standard APIs too though). The MTT Chunxiao GPU is clocked at 1.80 GHz – 1.90 GHz and packs 4,096 stream processors, 128 tensor cores, 256 texture units, and 256 render output. The GPU features a 256-bit memory interface and can work with GDDR6 memory with a 14 GT/s data transfer rate. As for host bus, it features 16 PCIe Gen5 lanes and fully supports the SR-IOV specification for PCIe virtualization in server environments. The GPU up to 32-way partitioning, which will be handy for Android games rendering.  

The chip also comes with a rather powerful video engine that supports AV1, H.264, and H.265 codecs for up to 8K videos as well as can decode up to 32 streams at a 1080p30 resolution. As for output, the GPU has four display engines supporting an up to 8Kp30 resolution and has three DisplayPort 1.4 as well as one HDMI 2.1 interface. 

Moore Threads says that depending on clocks, its Chunxiao GPU can deliver 14.4 FP32 TFLOPS or 15.2 FP32 TFLOPS, which is comparable to single precision compute performance of Nvidia's GeForce RTX 3060 Ti GPU. On paper, this is enough to make it to our list of the best gaming graphics cards available today, but how the ChunxiaoGPU works in real life is something that we will have to test. 

Meanwhile, since the Moore Threads Chunxiao graphics processor is aimed at AI, gaming and datacenter workloads, it comprises of 22 billion of transistors, which is fairly complex. To put the number into context, Nvidia's GA104 features 17.4 billion transistors, whereas AMD's Navi 21 packs 26.8 billion of transistors. Moore Threads does not disclose which process technology it uses to make its Chunxiao graphics processor.

Two Products Coming 

For now, Moore Threads plans to offer two products based on the Chunxiao GPU — the MTT S80 graphics card with a 14.4 FP32 TFLOPS throughput and 16GB of memory as well as the MTT S3000 server card with a 15.2 FP32 TFLOPS compute performance and 32GB of memory. 

Moore Threads has not disclosed power consumption of its MTT S80 and MTT S3000 products, though the former comes with a rather sophisticated cooling system with three fans. 

One of the main advantages that the Moore Threads Chunxiao offers is its broad compatibility; It can work with a variety of client and datacenter hardware platforms (Arm, Ampere, Intel, etc.) and operating systems. It is compatible with Microsoft's DirectX (and therefore will ship with appropriate drivers for Windows), Khronos Group's OpenGL/OpenGL ES, proprietary MUSA as well as multiple specialized APIs. Furthermore, the Chunxiao can work with PyTorch, TensorFlow, PaddlePaddle, Jittor and other mainstream deep learning frameworks and popular AI models. 

To ensure compatibility with modern games, Moore Threads says it had worked closely with developers of Unreal Engine and Unity as well as creators of popular titles (e.g., Call of Duty, Crossfire, Counter Strike, Diablo 3, League of Legends, etc.). Meanwhile, Moore Threads admits that for now its Chunxiao only properly supports about 20 DirectX titles and does not make any promises about performance. 

The GPU developer plans to start selling its MTT S80 graphics cards at JD.com on November 11, 2022.  No pricing has been announced.

Chinese GPU start-up Moore Threads has taken the very significant step of shipping its first graphics cards, according to a report published by PC Watch. The fledgling MTT S10 graphics card is listed on Taobao, a Chinese retail site for 800 Chinese Yuan, noted the source, but didn’t provide any seller links. Moore Threads is a particularly interesting Chinese “unicorn” start-up, as it is well funded and aims to market “fully featured” domestic GPUs. Moreover, its core engineering team is purportedly laden with ex-Nvidia talent. Conversely, Chinese GPU rival MetaX Tech is said to be majority staffed with engineers who have previous experience at AMD.

Moore Threads isn’t afraid of announcing products, but so far this MTT S10 graphics card appears to be the first to actually hit retail. Back in March, Moore Threads announced the MTT S60 for PC desktop and the MTT S2000 for server markets. Neither of those products have been seen since, but now the little MTT S10 is here we could shortly see some independent third party views on the abilities and performance of the underlying MUSA graphics architecture.

The Moore Threads MTT S10 is an admittedly low-power discrete GPU and highly unlikely to make our list of best gaming GPUs. According to the official website’s product pages, the MTT S10 targets “digital office” workloads. It is small in size, and low power consumption. Nevertheless it still supports multiple 3D graphics APIs such as OpenGL, OpenGL ES and Vulkan – as well as accelerating online WebGL. It has 2D appeal too, with support for HDMI 2.1 output to 4K 60HZ monitors, accelerated support for AV1, H.265, and H.264 video codecs and video conferencing.

When the Moore Threads MTT S60 and MTT S200 launched we managed to dig up some key tech specs but we aren’t so lucky with the perhaps low-key launch of this entry-level product. All we know about the MTT S10 is that it: uses the 12nm MUSA architecture for wide 3D, 2D, and codec support; it has a TDP of just 30W; the GPU clock is 1 GHz; it uses LPDDR4 memory; has a PCIe 3.0 x8 interface; and the half-height board pictured comes with just one HDMI 2.1 port and one D-Sub VGA connector. Sadly, we can’t find the MTT S10’s quota of GPU cores, estimated TFLOPS performance, nor even its VRAM amount.

Remember, the MTT S60 is being heralded by Moore Threads as an all rounder, capable enough for content creation and enjoying PC gaming. It was demonstrated at launch playing League of Legends with various enhanced graphics features and filters applied. The newly released MTT S10 might only offer half the GPU cores and VRAM though, we don’t know at the time of writing. We will be watching Chinese sources and social media for further information and any tests of More Threads first available GPU.

If you unearth the MTT S10 on Taobao and it ships to your location, it reportedly costs 800 Chinese Yuan, equivalent to approximately $112 or £100.

Moore Threads was founded in October 2020 and broke cover in late 2021 with the announcement that it would become China's first "fully featured GPU" company. Today the Beijing startup officially announced its first two graphics cards; the MTT S60 for PC desktops and workstations, and the MTT S2000 for servers. Both are based on 12nm GPUs built using the MUSA architecture, and the MTT S60 was demo'd at the event playing League of Legends at 1080p.

We wrote about Moore Threads when it exited stealth mode back in November of last year. Today at the launch event, founder and CEO of Moore Threads, Zhang Jianzhong, made his first public appearance in his current role. ITHome explains that Jianzhong is a former global VP of Nvidia and GM in China, and has been deeply involved in the graphics business for 15 years. He has a strong supporting team with experience across the industry in China and abroad.

Before looking closer at what was launched today, let us chart out the key hardware specs, as far as we know them at this early stage:

If you want a point of reference for the MTT S60's 6 TFLOPs GPU performance claim, the Xbox One X was heralded as having a 6 TFLOPs GPU back in 2017. The Nvidia GeForce GTX 1070 from a year earlier had pretty similar performance by this metric, of approximately 6.5 TFLOPs.

The first gen MUSA architecture has been dubbed 'Sudi'. According to reports from the launch event, this GPU features four major processing engines: graphics, AI, video, and physics.

As part of the launch presentation, there were a wide range of demos shown. For the demo of the graphics rendering engine, Moore Threads showed a game of League of Legends being played on an MTT S60. Not the most demanding game, but gamers will be happy to know that advanced features such as global illumination, spatiotemporal anti-aliasing, physical rendering, soft shadows, reflections, and volumetric light are said to be supported. The title ran at 1080p, but we don't know anything about the graphics quality or average framerate. One must also remember that LOL's official recommended specs are very low – to play fluidly, Riot only recommends an Nvidia GeForce 560 or AMD Radeon HD 6950 or better (updated specs from July 2021). 

Moving onto AI, the MUSA-based graphics cards can support various mainstream AI frameworks, such as those for accelerating visual processing, audio processing, natural language processing and more.

Video processing happens with what is called an intelligent multimedia engine. It is claimed to be capable of 8K codec support and is AV1 codec friendly for cloud conferencing, live broadcast and more.

Last, but not least, the physical engine, dubbed Alphacore, works with tools such as Unity, Houdini, and Unreal to accelerate realistic motion of complex structures and materials.

Demonstrations of multiple MUSA architecture GPU processing technologies working together were made in 3D rendering programs, image recognition, metaverse avatar animations, scientific simulation software and more.

In a Q&A, the CEO of Moore Treads echoed the wisdom that graphics card driver and support software is extremely important to a successful product. From Jianzhong's claims, it sounds like his company has everything covered in this department too. For example, the MUSA architecture is said have support for OpenCL, SYCL, CUDA, Vulkan, DirectX, OpenGL / GLES and other mainstream programming interfaces. Additionally, these GPUs will work with systems based on either x86 and Arm processors and all domestic (China) mainstream OSes.

It must also be commented upon that the progress shown by Moore Threads is extraordinary. From its founding in Oct 2020, it became a 'Unicorn' company within 100 days, and as per our report last November had successfully developed a fully featured GPU, with a multitude of modern supporting features and technologies, that has now finally been announced today in two SKUs.

Without pouring cold water on the achievements thus far, it is quite easy to make big claims about performance, capabilities, and compatibility. Moore Threads has had a lot of investment thrown at it and names multiple Chinese PC partners including Lenovo and Tencent, so it's not like it's a scrappy underdog. However, the demos were not extremely illuminating, and the proof of the product is when it is in end-user hands.

With the accelerated growth and development seen so far, it shouldn't take the outfit many more months to ship product to independent testers/the public, and then we can really see what the MTT S60 can achieve.

Graphics card prices have been easing significantly in recent weeks. If you don't want to wait for a Moore Threads MTT S60 to become available, you could instead look over our Best Graphics Cards for Gaming in 2022 guide.

A Chinese company called Moore Threads has recently announced that it has become the first national company to introduce a domestic, "fully featured" GPU solution. The company reportedly tapped most of its core engineering team from Nvidia, with the usual suspects (Microsoft, Intel, Arm, and others) also being headhunted for the engineering talent that makes up the company's roster. "Moore Threads" is more than an interesting naming choice - the company's website does promise "double the number of concurrent threads every two years."

If that sounds familiar, it's because it refers to Moore's Law, which predicted that the number of transistors in an integrated circuit would double every two years. Moore's Law is named for Intel co-founder Gordon Moore.

The company claims that its GPU solutions are based on fully domestic intellectual property and manufacturing expertise, and are capable of a variety of GPU workloads, including 3D graphics computing, AI training and inference computing, high-performance parallel computing, and acceleration for ultra-high-definition video codecs. The company aims for its GPUs to be seamlessly integrated with partner companies' infrastructure designs, including Chinese-developed CPU and platform solutions — an integration process that the company has already started. Moore Threads claims it already has accumulated a wealth of knowledge for every stage of GPU production, and that that's the enabler for its claim of being able to design, manufacture, market and service its GPUs, covering the entire product life cycle.

Interestingly, Moore Threads was founded recently, in October 2020, so a mere year has come and pass since its inception and march toward the development of this GPU solution. The company did enjoy ample funding for its endeavors, however. The startup has already gone through three funding rounds in a single year, with prominent investors such as Sequoia Capital China, ByteDance (of TikTok fame) and Tencent all reportedly contributing at one point or another. The latest Series A funding round  brought the company some $313 million, with the aim to enable Moore Threads to kickstart the mass production and manufacturing of its first GPU chip and continued IP research and development for its GPU SoCs.

If Moore Threads' claims ring true, then this is certainly one of the pivotal moments in China's plans to become technologically independent from Western countries when it comes to silicon-bound technologies. With pressure pilling on China's manufacturing capabilities in the face of the U.S.–China trade war, technological independence for China does seem to be dependent on that country's ability to reinvent the wheel.