Huawei backs HBM memory manufacturing in China to sidestep crippling US sanctions that restrict AI development

News
By
published

If you want AI and HPC, you must have HBM.

AMD
(Image credit: AMD)

Huawei is reportedly building a consortium of memory producers to build high bandwidth memory (HBM) to break free of Western sanctions. HBM is crucially important for AI and HPC processors because no matter how much compute power they have, performance is often limited by memory bandwidth. Huawei understands this perfectly, so it is now backing the development of HBM2 memory at one China-based DRAM maker, according to The Information. 

The Huawei-led consortium is reportedly backed by the Chinese government and includes several other Chinese semiconductor companies, including memory maker Fujian Jinhua Integrated Circuit, as well as specialists in advanced chip packaging. The consortium is currently developing HBM2 memory, which is generations behind what market leaders often use. The facilities even have two production lines. Huawei reportedly aims to complete development and start mass production of HBM2 memory by 2026. 

In addition, ChangXin Memory Technologies (CXMT) is also trying to develop HBM technology, though its achievements in this realm are unclear. There is also a media report that XMC, another China-based semiconductor company, is also working on an HBM project. XMC is controlled by Yangtze Memory Technology Co. (YMTC), China's leading producer of 3D NAND, which is in turn controlled by state-owned Tsinghua Unigroup. 

HBM is currently sold by Micron, Samsung, and SK hynix, and these memory stacks are not available on the spot market. Since HBM memory is made using American technologies, these companies have to apply for an export license from the U.S. government, which reviews the applications with a presumption of denial.  

Huawei badly needs HBM for its Ascend-series processors for AI applications, and while SMIC (presumably) can make these chips for Huawei, HBM availability is clearly a bottleneck for Huawei's AI processors. It is unclear how the company gets HBM memory now, though it is theoretically possible to get these stacks indirectly. Of course, this means limited availability for Huawei and leads to limited availability of Ascend processors as the Chinese high-tech giant primarily wants to use these devices for its own AI services. 

Of course, Huawei's and other HBM companies' endeavors face considerable obstacles, notably the international regulations that limit sales of advanced chipmaking equipment to China.

Huawei's formation of the group and CXMT's efforts can be considered another strategic move by China to become self-sufficient in key technologies for AI and supercomputing.

Anton Shilov
Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

7 Comments Comment from the forums
  • usertests
    It would be amazing if China started increasing the global HBM supply and the AI bubble subsequently burst. Let's get HBM in consumer GPUs/CPUs/APUs already (VRAM, L4 cache).
    Reply
  • edzieba
    usertests said:
    It would be amazing if China started increasing the global HBM supply and the AI bubble subsequently burst. Let's get HBM in consumer GPUs/CPUs/APUs already (VRAM, L4 cache).
    Even if bit-for-bit HBM was at price-parity with DRAM, cards using it would be substantially more expensive due to the need for the silicon interposer.

    Now, if China eschewed HBM and revived HMC...
    Reply
  • Lewinator56
    edzieba said:
    Even if bit-for-bit HBM was at price-parity with DRAM, cards using it would be substantially more expensive due to the need for the silicon interposer.

    Now, if China eschewed HBM and revived HMC...
    May I remind you of the R9 fury series and vega.

    HBM works on consumer GPUs, but HBM2 had it's limitations. GDDR6 just made more sense as it is so much easier and cheaper to work with.
    Reply
  • edzieba
    Lewinator56 said:
    May I remind you of the R9 fury series and vega.
    Very overpriced and underperformed?

    And also acted as a preview of the extremely variable power delivery demands that have now become commonplace amongst GPUs - and was a surprise wakeup call for SFX PSU makers, where the R9 Nano found a niche but was nearly unusable due to it tripping PSU overcurrent protection circuits under load, with PSUs that met the specsheet power draw provided by AMD, but could not meet the transient current draw the card actually demanded. The Vega/Fury series is why GPUs started vastly overstating power supply requirements - to handle those transient current spikes - even after PSU manufacturers tweaked designs to tolerate those.
    Reply
  • Lewinator56
    edzieba said:
    Very overpriced and underperformed?

    And also acted as a preview of the extremely variable power delivery demands that have now become commonplace amongst GPUs - and was a surprise wakeup call for SFX PSU makers, where the R9 Nano found a niche but was nearly unusable due to it tripping PSU overcurrent protection circuits under load, with PSUs that met the specsheet power draw provided by AMD, but could not meet the transient current draw the card actually demanded. The Vega/Fury series is why GPUs started vastly overstating power supply requirements - to handle those transient current spikes - even after PSU manufacturers tweaked designs to tolerate those.
    The R9 fury series didn't perform poorly, but they weren't significantly faster than their predecessors, they were also quite expensive and used first generation HBM.

    The Vega series however performed extremely well, the Vega64 released as a contender (a yearish later) for the GTX1080, and did particularly well at that, maintaining an average 5-10% lead in performance, while also costing less.

    Transient power spikes won't come from the memory though, rather the core design, GCN was really on its way out with the R9 fury and HBM was I guess an attempt to try to get every little bit of performance out of it as possible. Polaris, while still GCN, was a massive step, and Vega built on that. AMD had pretty much been using the same core from the HD 7000 series with tweaks up until Polaris.
    Reply
  • edzieba
    Lewinator56 said:
    The Vega series however performed extremely well, the Vega64 released as a contender (a yearish later) for the GTX1080, and did particularly well at that, maintaining an average 5-10% lead in performance, while also costing less.
    'Launch MSRP' matched the GTX1080, but launch MSRP was not available for long (the first few hundred cards per retailer) with real-world pricing rapidly rising above the GTX1080.
    Performance was for all intents and purposes matched with the GTX1080 (some games a few percent faster, some slower) but it was clear from the power consumption being ~1.5x the GTX1080 that the cards were run at the highest voltage AMD could release with to push the clocks necessary to meet that performance goal. This also put the Vega series in the unique situation where any meaningful overclock required undervolting to open up any meaningful power-limit headroom - as well as a healthy dose of silicon lottery to have a card that could operate at even stock clocks when undervolted.
    Reply
  • Lewinator56
    edzieba said:
    'Launch MSRP' matched the GTX1080, but launch MSRP was not available for long (the first few hundred cards per retailer) with real-world pricing rapidly rising above the GTX1080.
    Performance was for all intents and purposes matched with the GTX1080 (some games a few percent faster, some slower) but it was clear from the power consumption being ~1.5x the GTX1080 that the cards were run at the highest voltage AMD could release with to push the clocks necessary to meet that performance goal. This also put the Vega series in the unique situation where any meaningful overclock required undervolting to open up any meaningful power-limit headroom - as well as a healthy dose of silicon lottery to have a card that could operate at even stock clocks when undervolted.
    Luckily, I wasn't exposed to the launch pricing. I got my V64 about a year after launch for £399 for a nitro+, which was a pretty damn good deal if you ask me. But launch pricing went a bit wacky.

    The power consumption is an interesting one, I never really fiddled with the power limit on my card, left it at default which I think was 250W, never really gained much from increasing it but I think I only went as high as 300W, but I was running ~50MHz off the boost clocks at the stock limit, so maybe I won the silicon lottery? (With a nitro+ I guess that makes sense). Was never thermally limited either. The only thing mine didn't like was a memory overclock. I think I got 30MHz over the base before it became unstable.
    Reply