Huawei said to have built an AI GPU that matches Nvidia's A100

[midian182]

In a nutshell: Could Nvidia's dominance in the AI and high-performance computing hardware segment in China face a challenge from Huawei? According to the chairman of a Chinese AI company, the sanctioned tech giant now has a GPU that offers similar performance to Nvidia's A100 GPU.

Liu Qingfeng, the founder, chairman, and president of Chinese AI firm iFlytek, which supplies voice recognition software, made the claim at the 19th Summer Summit of the 2023 Yabuli China Entrepreneurs Forum (via IT Home).

Qingfeng said that Huawei has been making massive strides when it comes to GPUs, having reached performance parity with Nvidia's A100.

Tens of thousands of Team Green's A100 GPUs, which cost around $10,000 each, were used in a supercomputer to train ChatGPT. Only the Hopper H100 sits above it in Nvidia's AI product stack.

A domestic company producing GPUs equivalent to the A100 would be a major step for China. US sanctions prevent Nvidia from selling the A100 to firms in the country, leading Nvidia to create a cut-down version for China in the form of the A800, which has an interconnect speed of 400 GB/s, down from the A100's 600 GB/s. The scarcity of the A100 has resulted in a black market for Nvidia AI chips in the Asian nation, where they sell for double their MSRP and without warranties.

Qingfeng added that three Huawei board members are working with a special team from his company. He said that there are no problems with algorithms in the field of AI in China, but the computing power is held down by Nvidia.

Elsewhere, Qingfeng said that iFlytek will release a new general-purpose AI model to compete with ChatGPT. It will launch in both English and Chinese versions, and while it won't be on the same level as OpenAI's product at first, the company says it should be able to match the GPT-4 LLM by the first half of next year.

iFlytek officially released its Spark cognitive large-scale model in May. It has seven core capabilities: text generation, language understanding, knowledge-based questioning and answering, logical reasoning, mathematical ability, code ability, and multimodal capability. The company also announced a partnership with Huawei this month to design all-in-one machines for AI.

Permalink to story.

https://www.techspot.com/news/99948-huawei-alleged-have-ai-gpu-matches-nvidia-a100.html

 

[winjer]

I really doubt this is true.

 

[Neatfeatguy]

I really doubt this is true.

But, it's as true as how China says they have built a quad core CPU that nearly matches first gen Ryzen.

https://www.techspot.com/news/90544-china-new-loongson-cpu-almost-fast-first-ryzen.html

So, if they say it's true....well, you better believe them because of reasons.

 

[neeyik]

I really doubt this is true.

To be fair to Huawei, it didn't say exactly what capability it had the same as the A100 -- it could well be just FP32 throughput, which isn't a strength of Nvidia's chip at all!

 

[winjer]

To be fair to Huawei, it didn't say exactly what capability it had the same as the A100 -- it could well be just FP32 throughput, which isn't a strength of Nvidia's chip at all!

If they are referring to AI, then it's about matrix operations. Most of them in lower percision than FP32.

 

[Uncle Al]

Doesn't matter, I still don't trust them one bit .....

 

[neeyik]

If they are referring to AI, then it's about matrix operations. Most of them in lower percision than FP32.

Oh, indeed -- I was just making a bit of a (rubbish) joke, as one can easily claim to be an opponent if one simply targets its weakest attribute.

The GA100 is a wall of tensor cores and cache, with the former being in its third revision. A well-designed ASIC could, in theory, hold up just as well, with a good example being Google's TPU V4. The question is whether Huawei has managed to do the same.

 

[winjer]

Oh, indeed -- I was just making a bit of a (rubbish) joke, as one can easily claim to be an opponent if one simply targets its weakest attribute.

The GA100 is a wall of tensor cores and cache, with the former being in its third revision. A well-designed ASIC could, in theory, hold up just as well, with a good example being Google's TPU V4. The question is whether Huawei has managed to do the same.

My doubts are not just the necessary capabilities of developing such complex chips.
But also, that China no longer has access to TSMC most advanced process nodes.
AI chips also rely heavily on the memory subsystem, and I'm not sure if China has access to HBM3.

 

[neeyik]

AI chips also rely heavily on the memory subsystem, and I'm not sure if China has access to HBM3.

CMXT is definitely gaining ground in DRAM fabrication, but it's still a few years behind the likes of Micron and SK Hynix. Technically, there's nothing super advanced about HBM -- it is, after all, just stacked DRAM dies with power and data lines run through TSVs. So Huawei probably will have access to something like HBM or maybe HBM2, but when Samsung is stacked up to 12 layers of 10nm dies, CMXT has got its work really cut out.

 

[PEnnn]

The US never believed China would have a gigantic economy
Or a huge nuclear-powered navy
Or world-wide influence
Or a large air force with stealth fighters

And reading the comments here, the trend continues!

 

[neeyik]

The US never believed China would have a gigantic economy
Or a huge nuclear-powered navy
Or world-wide influence
Or a large air force with stealth fighters

And reading the comments here, the trend continues!

The doubt isn't about China, it's specifically focused on Huawei -- the company spends billions on R&D every year, and its product portfolio is enormous, but chip design and manufacturing hasn't been one of them. Even now, its latest phones use Qualcomm's SoCs. It's a big jump to go from using other firms' products to completely designing and fabricating a completely new ASIC that will directly compete against one that has years of experience and use behind it.

Edit: Well I certainly need to retract the above statement -- Huawei does design large-scale ASICs. Apologies for that!

 

[AIDailyDigital]

Huawei Enployee: "Hey boss, if we chain 139 of our crappy GPUs together in parallel and pump 91000 watts into it, we can match a single A100."

Boss: "Call state media and tell them of our new success! Take that NVIDIA!!!"