Having started its development in 2023, Loongson Technology's 9A1000 graphics card is one step closer to the finishing line. According to Chinese media outlet ITHome, the development of the 9A1000 has been completed, and the graphics card will begin tapeout in the third quarter of this year.
The 9A1000 is Loongson's first graphics card, marking a significant milestone for the Chinese manufacturer, which had previously focused mainly on processors. The company positions the 9A1000 as an entry-level graphics card that supports AI acceleration. Therefore, it doesn't compete in the same segment as the Lisuan G100, another Chinese graphics card, that allegedly rivals the GeForce RTX 4060.
Although we've been aware of the 9A1000 for some time, we still don't know its exact specifications. Loongson has kept details under wraps, only hinting that it offers performance similar to the Radeon RX 550, a card that was released eight years ago. However, it's too early to call it a win for Loongson, since the manufacturer still has to evaluate the 9A1000 after the tapeout.
The latest update on the 9A1000 indicates that Loongson has apparently reduced the area of the stream processor by 20%. The manufacturer also claims that it has increased the 9A1000's operating frequency by 25%, while optimizing power consumption during light loads by 70%. In terms of feature set, the 9A1000 supports OpenGL 4.0 and OpenCL ES 3.2 APIs.
According to Looongson, the 9A1000 is up to 4X faster than the LG200, the integrated graphics unit inside the 2K3000 processor. The 9A1000 also provides up to 40 TOPS of AI computing power, which is slightly below that of AMD's XDNA 2 NPU inside the Ryzen AI Max+ (codenamed Strix Halo) chips, which deliver up to 50 TOPS.
The 9A1000 isn't the only graphics card on Loongson's plate. The company is also working on the 9A2000, which it claims is up to 10X faster than the 9A1000, with performance levels comparable to those of the GeForce RTX 2080. There are also plans for a 9A3000, a follow-up to the 9A2000, but no known specifications are available yet.
Although it may not be immediately apparent, numerous Chinese corporations and startups have entered the graphics card industry. However, many of these entities fail and subsequently cease operations. Therefore, we only hear news about the more prominent firms, such as Biren, Moore Threads, and, in recent years, Loongson and Lisuan Technology.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button!
Zhiye Liu is a news editor, memory reviewer, and SSD tester at Tom’s Hardware. Although he loves everything that’s hardware, he has a soft spot for CPUs, GPUs, and RAM.
-
JarredWaltonGPU I keep waiting for China's GPU companies to do something really interesting. I know gaming GPUs are hard, since you need both capable hardware and capable drivers/software. Still, RX 550 was a slug-slow junk GPU that was barely worth mentioning way back in 2017 when it first arrived.Reply
Even the RX 560 was basically trash hardware, and the 550 was like, "Watch this as we cut the compute in half!" It was a $79 MSRP GPU in 2017, and no one wanted it. How much is the 9A1000 going to cost in China? Probably more than the launch 550 price, with worse drivers — and maybe worse power and efficiency as well!
Again, I know it's difficult to make a good, new GPU architecture. (Just ask Intel...) But at least Intel managed to basically match the RTX 3060 with Arc A750, and it was only two and a half years later. China needs to do much more than matching eight years old budget GPUs. -
TCA_ChinChin Reply
I mean this will be in the same tier as gt710s and bygone old and slow GPUs used to add extra monitors and a modicum of compute to old platforms. Its a trash GPU but I dont see how this is indicative of the whole Chinese GPU industry. Loongson from what i know is more focused on getting their own CPUs off the ground. Im surprised they made this GPU industry the first place. If you're looking for the bleeding edge of GPU performance in China, Moores Threads and Lisuan (which are mentioned in this article) are probably a better indicator. Still not great compared to what Intel, AMD and Nvidia have, but improving greatly.JarredWaltonGPU said:I keep waiting for China's GPU companies to do something really interesting. I know gaming GPUs are hard, since you need both capable hardware and capable drivers/software. Still, RX 550 was a slug-slow junk GPU that was barely worth mentioning way back in 2017 when it first arrived.
Even the RX 560 was basically trash hardware, and the 550 was like, "Watch this as we cut the compute in half!" It was a $79 MSRP GPU in 2017, and no one wanted it. How much is the 9A1000 going to cost in China? Probably more than the launch 550 price, with worse drivers — and maybe worse power and efficiency as well!
Again, I know it's difficult to make a good, new GPU architecture. (Just ask Intel...) But at least Intel managed to basically match the RTX 3060 with Arc A750, and it was only two and a half years later. China needs to do much more than matching eight years old budget GPUs.
I dont find it unusual that a world class chip company that already had years of prior experience with iGPUs and compute like Intel makes significantly better GPUs than 5 year old startups and a mid-at-best CPU company that started from MIPs. -
das_stig Yeh the 550 sucked, but so did the intel HD series and that powered how many generations of laptops?Reply
It's good enough and has the AI tops, then for business what more do you need and it's home grown?
I'm no fan of China, but people need to stop comparing their GPUs targets with our and be grown up and not Trumpies Children ! -
das_stig Reply
Intel has had decades and many false starts, China has had few years, another bias comment.JarredWaltonGPU said:I keep waiting for China's GPU companies to do something really interesting. I know gaming GPUs are hard, since you need both capable hardware and capable drivers/software. Still, RX 550 was a slug-slow junk GPU that was barely worth mentioning way back in 2017 when it first arrived.
Even the RX 560 was basically trash hardware, and the 550 was like, "Watch this as we cut the compute in half!" It was a $79 MSRP GPU in 2017, and no one wanted it. How much is the 9A1000 going to cost in China? Probably more than the launch 550 price, with worse drivers — and maybe worse power and efficiency as well!
Again, I know it's difficult to make a good, new GPU architecture. (Just ask Intel...) But at least Intel managed to basically match the RTX 3060 with Arc A750, and it was only two and a half years later. China needs to do much more than matching eight years old budget GPUs. -
Notton While I'm not desperate for a GPU right now, it would be nice to have a cheap GPU to go with my 5800X3D. Cheaper than a 9060XT hasn't been an option for a while now because they don't exist.Reply -
zsydeepsky ReplyJarredWaltonGPU said:I keep waiting for China's GPU companies to do something really interesting. I know gaming GPUs are hard, since you need both capable hardware and capable drivers/software. Still, RX 550 was a slug-slow junk GPU that was barely worth mentioning way back in 2017 when it first arrived.
You probably expected the wrong company. Loongson is a NEW PLAYER among Chinese GPU makers; their expertise was mostly on CPU design.
Therefore, this RX-550 level card has basically 2 goals:
- better performance than Loongson's integrated GPU (which, honestly speaking, is just garbage)
- tech verification for their GPGPU architecture
I listened to their introduction, so this 9A1000 is like a verification of a (big) CUDA core of Nvidia cards. They plan to use 9A1000 to verify the architecture works, then later make 9A2000 card, which, according to their plan, basically is 9A1000 x4 (per chip) x2 (dual chip), which will be at ~2080 level performance. All these will be done in mature nodes (12-14nm, I presume). If 9A2000 also went well, then they will push the design to a more advanced node (5/7 nm, I presume), which Loongson claims will provide A100/H100 level performance.
It's like a reversed process of how RX-550 was created from the flagship card. You make RX-550 first, then doubles it to create RX-580 or even AMD Radeon Pro Duo.
They are still in the prototyping stage, so I wouldn't expect anything spectacular to come from them before 2030. -
bit_user Reply
I have a 4 GB model in my fileserver. It's great for driving dual monitors @ 4k and can run Google Earth effortlessly (my standard "benchmark" is Downtown London).JarredWaltonGPU said:RX 550 was a slug-slow junk GPU that was barely worth mentioning way back in 2017 when it first arrived.
LOL, try finding any new graphics card for that, today!JarredWaltonGPU said:It was a $79 MSRP GPU in 2017, and no one wanted it.
: )
Maybe, but the RX 550 was made on GF 14 nm. And Loongson knows a little about chip design, so I think it might not actually be so bad on the efficiency front.JarredWaltonGPU said:and maybe worse power and efficiency as well!
(edit: just saw @zsydeepsky 's post. So, maybe not great on the efficiency front, if it's using basically the same node) -
bit_user Reply
What sort of performance level are you targeting? Intel's A380 (or the Arc Pro A40) would be good entry-level options that can drive multiple monitors at a decent rate and have fairly comprehensive codec acceleration.Notton said:While I'm not desperate for a GPU right now, it would be nice to have a cheap GPU to go with my 5800X3D. Cheaper than a 9060XT hasn't been an option for a while now because they don't exist.
If you only need dual monitor support, you could drop down to a Radeon Pro W6400, but you lose AV1 decode acceleration and most encode acceleration. However, it's single-slot, half-height and low-profile (in case that's important). -
Notton Reply
It needs AV1 decode, preferably encode.bit_user said:What sort of performance level are you targeting? Intel's A380 (or the Arc Pro A40) would be good entry-level options that can drive multiple monitors at a decent rate and have fairly comprehensive codec acceleration.
If you only need dual monitor support, you could drop down to a Radeon Pro W6400, but you lose AV1 decode acceleration and most encode acceleration. However, it's single-slot, half-height and low-profile (in case that's important).
I would have gotten an Intel Arc A 310/380, except they have been overpriced with zero stock everywhere I look.
The only option is the RTX3050 6GB because the other choices are GT1030 and GT710/730, and I sure as heck ain't buying another nvidia card. -
JarredWaltonGPU Reply
I know it's not indicative of all of China's GPU efforts, but this is very, VERY low down the totem pole. We're talking about a new potentially dedicated GPU that probably sits about on par with the Arc A380. Again, that's four years old now.TCA_ChinChin said:I mean this will be in the same tier as gt710s and bygone old and slow GPUs used to add extra monitors and a modicum of compute to old platforms. Its a trash GPU but I dont see how this is indicative of the whole Chinese GPU industry. Loongson from what i know is more focused on getting their own CPUs off the ground. Im surprised they made this GPU industry the first place. If you're looking for the bleeding edge of GPU performance in China, Moores Threads and Lisuan (which are mentioned in this article) are probably a better indicator. Still not great compared to what Intel, AMD and Nvidia have, but improving greatly.
I dont find it unusual that a world class chip company that already had years of prior experience with iGPUs and compute like Intel makes significantly better GPUs than 5 year old startups and a mid-at-best CPU company that started from MIPs.
Yes, you can double, then double again, and again, etc. and use this as a building block. But if this is the core building block, what are we talking about? RX 550 levels of performance would be something like maybe a current two SM Nvidia design, or two CU AMD design. That's super tiny. All the other stuff for even a 32-bit memory interface would be more than a dual-CU/SM chip might need.zsydeepsky said:You probably expected the wrong company. Loongson is a NEW PLAYER among Chinese GPU makers; their expertise was mostly on CPU design.
Therefore, this RX-550 level card has basically 2 goals:
- better performance than Loongson's integrated GPU (which, honestly speaking, is just garbage)
- tech verification for their GPGPU architecture
I listened to their introduction, so this 9A1000 is like a verification of a (big) CUDA core of Nvidia cards. They plan to use 9A1000 to verify the architecture works, then later make 9A2000 card, which, according to their plan, basically is 9A1000 x4 (per chip) x2 (dual chip), which will be at ~2080 level performance. All these will be done in mature nodes (12-14nm, I presume). If 9A2000 also went well, then they will push the design to a more advanced node (5/7 nm, I presume), which Loongson claims will provide A100/H100 level performance.
It's like a reversed process of how RX-550 was created from the flagship card. You make RX-550 first, then doubles it to create RX-580 or even AMD Radeon Pro Duo.
They are still in the prototyping stage, so I wouldn't expect anything spectacular to come from them before 2030.
Normally, the base building unit would be at least 8 SM/CU. AMD's integrated graphics in Zen 4/Zen 5 desktop are only 2 CUs, but unless I read it wrong, this is a dedicated GPU. Basically, one quarter the performance of what I'd think of as a reasonable baseline for 2025, for a new entry into the GPU space.