News Huawei's GPU Reportedly Matches Nvidia's A100: Report

[Admin]

Founder of iFlytek claims that Huawei's compute GPU can match Nvidia's A100 performance.

Huawei's GPU Reportedly Matches Nvidia's A100: Report : Read more

 

[Nikolay Mihaylov]

Is it a GPU or an AI accelerator?

 

[atomicWAR]

Is it a GPU or an AI accelerator?

Sounds like its a bit of both just like Nvidia's offerings. Question is can play Crysis? /s

 

[Kamen Rider Blade]

Question is, how much of nVIDIA hardware IP did they steal?

 

[P1nky]

A report that reportedly reports.

 

[atomicWAR]

Question is, how much of nVIDIA hardware IP did they steal?

Most if not all if China's recent tech history is any indiction (not being political, just facts). Though we do need more competition in the gpu space, stealing is not the way. So even if they released a gaming card...between spying concerns and IP theft, no thank you.

A report that reportedly reports.

And I thought it was a report the reportedly reports a reporting. My bad thanks for clarifing the issue.

 

[gg83]

I don't believe they have a competitive anything. Unless it's an a100 with a huawei sticker.

 

[bit_user]

Is it a GPU or an AI accelerator?

Given the article says:

"the head of iFlytek specifically noted that he was talking about a compute GPU"
​

I'd assume it's a GPU-like compute accelerator with no graphics capabilities.

 

[bit_user]

I don't believe they have a competitive anything. Unless it's an a100 with a huawei sticker.

Yeah, I really don't trust anything that's not from an independent reviewer.

And that goes for not just the hardware, but also talk of AI models:

​

"Back in May, iFlytek officially launched a cognitive large-scale model featuring seven core abilities: text generation, language understanding, knowledge-based question-answering, logical reasoning, mathematical ability, code ability, and multi-modal capability."​

​

It might have those capabilities, in theory, but we need to know if those features are actually usable and how they compare with GPT-4 and other LLMs.

 

[fleurdelis]

They never claimed it was a GPU, they said it's an AI accelerator. So ... just like Google's TPUs, which have been
competitive with Nvidia for 7 years now.

Whether the press release actually matches reality is one thing, but the comments in here saying that China is not capable of doing this without IP theft is just wishful thinking. It's an embarrassingly parallel problem space that boils down lots of parallel FMA/FMAC units.

The harder part is the software compatibility (esp pytorch & tensorflow). Google getting this right is why TPUs have been viable. Without that you might have a supercomputer but you don't have something that's going to gain mass adoption (example: AMD's competitiveness in the HPC space but complete failure in AI).

 

[bit_user]

Whether the press release actually matches reality is one thing, but the comments in here saying that China is not capable of doing this without IP theft is just wishful thinking. It's an embarrassingly parallel problem space that boils down lots of parallel FMA/FMAC units.

Ah, the hubris of inexperience. If it were that simple, Nvidia wouldn't be so dominant.

You're in plenty of company, though. I'm sure there have been hundreds of AI ASICs and IP blocks designed by now, the vast majority of which have fallen by the wayside, once they met up against the harsh realities of what it takes to be truly competitive.

It's one thing to make a toy IP for doing inference on tiny CNN models. It's a completely different universe, when you look at the challenges of training multi-billion parameter complex models.

The harder part is the software compatibility (esp pytorch & tensorflow). Google getting this right is why TPUs have been viable. Without that you might have a supercomputer but you don't have something that's going to gain mass adoption (example: AMD's competitiveness in the HPC space but complete failure in AI).

It's funny of you to say that, because a lot of AI chipmakers have long claimed such compatibility, including AMD. As for Google, you know they created TensorFlow, right?

 

[nookoool]

Question is, how much of nVIDIA hardware IP did they steal?

IMO If you will neither let them buy or allow them to manufacturer product A, IP is out the window for that particular product.

 

[bit_user]

IMO If you will neither let them buy or allow them to manufacturer product A, IP is out the window for that particular product.

Speaking hypothetically about "IP", in general: it's not as if the only thing you could do with it is to make exact, off-brand replicas. There are lots of ideas, algorithms, and techniques embbeded within it, that could be copied.

Even at the RTL level or below, Nvidia first introduced Tensor cores on the 12 nm process node, which is presumably a node they still have access to.

 

[Flayed]

I don't believe they have anything close to the performance they claim

 

[bit_user]

I don't believe they have anything close to the performance they claim

I think announcements like these are most often made for internal, often political reasons.

 

[Li Ken-un]

There are two ways this will turn out:

1. They languish, reinforcing the image of IP theft. Whether they’re using their own designs or not is irrelevant to the court of public opinion.
2. They leapfrog ahead of the West in the technology, thereby making it logically difficult to accuse them of IP theft; if they have advanced technology that others do not, then it follows that it could not have been stolen. It’s probably the only way China can ever shake its image as a nation of knockoffs and IP theft.

 

[USAFRet]

OK!

This has reached its inevitable end....insults and nationalistic jabs all around.

Relevant comments deleted, and thread closed.