China's Huawei Develops New AI Chip, Seeking To Match Nvidia (wsj.com) 52
Huawei is gearing up to test its newest and most powerful AI processor, which the company hopes could replace some higher-end products of U.S. chip giant Nvidia. From a WSJ report: Huawei has approached some Chinese tech companies about testing the technical feasibility of the new chip, called the Ascend 910D, people familiar with the matter said. The company is slated to receive the first batch of samples of the processor as soon as late May, some of the people said.
The development is still at an early stage, and a series of tests will be needed to assess the chip's performance and get it ready for customers, the people said. Huawei hopes that the latest iteration of its Ascend AI processors will be more powerful than Nvidia's H100, a popular chip used for AI training that was released in 2022, said one of the people. Previous versions are called 910B and 910C.
The development is still at an early stage, and a series of tests will be needed to assess the chip's performance and get it ready for customers, the people said. Huawei hopes that the latest iteration of its Ascend AI processors will be more powerful than Nvidia's H100, a popular chip used for AI training that was released in 2022, said one of the people. Previous versions are called 910B and 910C.
Re: No one cares (Score:1)
Re: (Score:2)
OTOH, it will probably be less efficient and generate more heat than equivalent NVIDIA products. China is reportedly still several years away from high quality production of large scale wafers.
(OTOH, with their current motivation they'll probably catch up in well under a decade.)
Re: (Score:1)
Nice try maggot, now go back to your la Presidenta shrine and perform another sacrifice....maybe offer up a migrant to the stormtroopers and ask for another gold star.
Re: (Score:2)
How quickly you'll sacrifice truth to make sure your narrative is preserved- that because of Trump, the US must just be the worst.
Ultimately, I think you're just as bad as that Orange cuck in the White House. In fact, I think you're the other half of that coin.
Re: (Score:2)
Re: (Score:2)
Most of the world lives in China or India, so your titled assertion is obviously incorrect. And this will also affect the market for everyone who imports from China, so most of the tech companies and their stockholders and investors will also be affected.
There are quite valid arguments that China is well behind the top "Western" companies. But they also appear to be catching up rather quickly. (Try to remember how Japan transitioned into a place for quality merchandise and high tech companies.)
Re: (Score:2)
Most of the world lives in China or India, so your titled assertion is obviously incorrect. And this will also affect the market for everyone who imports from China, so most of the tech companies and their stockholders and investors will also be affected.
There are quite valid arguments that China is well behind the top "Western" companies. But they also appear to be catching up rather quickly. (Try to remember how Japan transitioned into a place for quality merchandise and high tech companies.)
Except that in this case, it's not just Chinese companies that are struggling to catch up to Nvidia. AMD and Intel are struggling with their GPUs (especially aside from HPC sales). Google, Microsoft, Amazon, Meta, and others are struggling with their ASICs (aside from certain applications like recommendation systems and ads). Nvidia still has around 80% market share despite being more expensive than everyone else. And the Chinese companies are at least two generations behind in available fab processes (
Re: (Score:2)
It's my feeling that it will only be "suddenly" if you aren't paying attention to the in-between steps. E.g., after they can make good large wafers, they still need to draw patterns on them, and the machinery to do that comes from the Netherlands. If the EU were solidly in the US pocket, as it was a decade ago, then those machines couldn't be imported, but diplomatic relations have become a bit strained, so that's not clear.
Re: (Score:2)
Having a drunk on your staff is also a security risk. https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fapnews.com%2Farticle%2Fheg... [apnews.com]
Re: No one cares (Score:4, Funny)
Re: (Score:3)
No proven Huawei back doors? You've got an awful definition of "proven":
Finite Stateâ(TM)s tool checked 1.5 million firmware files from 558 Huawei enterprise networking products â" thatâ(TM)s just business systems, not consumer devices â" and found the average device had 102 vulnerabilities, at least a quarter of them severe enough to let a hacker get full access easily.
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fbreakingdefense.com%2F20... [breakingdefense.com]
Vodafone found hidden backdoors in Huawei equipment years ago, but company dragged its feet on fixing them
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Ffinancialpost.com%2Ftele... [financialpost.com]
And this data exfiltration wasnâ(TM)t a one-off operation. It had been occurring every night, at the same time, from January 2012 through January 2017. [....] Security experts swept the building and additionally discovered (and removed) bugs planted in desks and walls. (It was not specified to whom these bugs belonged, or if that was known, but the implication in the article seems to be linking this to the server spying.)
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fmedium.com%2Fdukeunivers... [medium.com]
Re: (Score:2)
That just seems like a lot of innuendo. They bought Huawei gear and somehow it installed bugs in their building, really? They crawled out at night and hid themselves in desks and walls?
The "backdoors" that Vodafone found where just the usual default management credentials stuff that Cisco and the like are well known for, and it was fixed. It's nothing like what we saw the NSA doing, systematically intercepting hardware being shipped and installing hardware rootkits.
It seems that China's government isn't stu
Re: (Score:2)
That story was very clear that China built the building for the AU, with the strong implication that the Chinese built the bugs into the building -- and coordinated the data exfiltration with whichever Chinese configured the servers and network gear. It was even more blatant and stupid than what you accuse the NSA of doing.
Not that a wumao shill like you would bother to read the source provided.
Re: No one cares (Score:2)
Ahh Americans, the rest of the world does exist
Tease Me With Products My Dictator Won't Allow (Score:2)
News you can't use!
Re: (Score:2)
The yields are so poor, that in a performance per dollar metric, NV is about 60% better, which is insane, because NV's pricing is predatory as fuck.
Re: (Score:2)
Fortunately for them, Trump has banned NVIDIA from competing with them. It's unfair Chinese companies should have to compete with American companies that have better products, so of course the American government is shielding them. Protectionism at its most insane and illogical!
Re: (Score:2)
Trump has banned NVIDIA from competing with them.
NV's export restrictions have existed on Democratic and Republican governments.
In fact, if there's one thing they seem to agree on, it's that they don't want the Chinese having our biggest and baddest number crunching parts.
Of course, this hasn't stopped them from smuggling them into the country.
Export controls are not protectionism.
They hurt NV, not help them.
NV would much prefer to sell their parts in China.
Re: (Score:2)
NV would much prefer to sell their parts in China.
Yes, that was my point. I think you missed the sarcasm in my post.
NV's export restrictions have existed on Democratic and Republican governments.
There were restrictions under Biden, which NVIDIA was able to work around with minor hardware changes. Trump's tariffs have completely closed the market to them. They aren't comparable.
Re: (Score:2)
Yes, that was my point. I think you missed the sarcasm in my post.
Unclear.
There were restrictions under Biden, which NVIDIA was able to work around with minor hardware changes. Trump's tariffs have completely closed the market to them. They aren't comparable.
You're right they're not comparable- because they're literally completely unrelated.
Export restrictions have nothing in common with import taxes.
You could argue that retaliatory tariffs from the Chinese (Chinese protectionism) would close the market to them, but our tariffs most definitely are not.
Just gonna steal the tech (Score:1)
It's the first class in every university in China, Corporate Espionage 101.
Re: (Score:2)
Cough.
Re: (Score:2)
Could be that the competition are restricted by decisions related to manufacturing node (aka CapEx wallet size), and NVIDIA's support of developers. Neither NVIDIA nor AMD has some advanced AI related GPU instructions that nobody else knows how to implement in hardware optimally. Nvidia H100 is manufactured on TSMC 4N, while AMD uses the cheaper 5 nanometer node for their MI300x. Therefore AMD had to make architectural decisions like increasing the chip size and number of transistors so that they can compet
Re: (Score:2)
Re: (Score:2)
Most of the performance differences for AI are coming from number of AI cores, local cache size & hierarchy, memory bandwidth .. basically balancing compromises of memory bound stuff is what is causing the major bottlenecks for AI. Those compromises exist due to the node physics. I'll go as far as saying two chip architects given the exact same node, die size, and application requirement (current generation AI inference or training).. would design very similar chips. It's not like AMD engineers didn't k
Re: (Score:2)
Most of the performance differences for AI are coming from number of AI cores, local cache size & hierarchy, memory bandwidth ..
Grossly simplified to the point of being wrong.
For inference- yes, the primary limiter is memory bandwidth. But that's simply because the memory bandwidth tends to be the bottleneck there. Cores are (practically) always capable of outpacing main RAM- otherwise you waste the utility of cache.
However, that does not mean that all cache hierarchies, clocks, and sharing strategies are created equal.
I'll go as far as saying two chip architects given the exact same node, die size, and application requirement (current generation AI inference or training).. would design very similar chips.
And this is just wrong.
From a microarchitectural perspective, the H200 simply eats the MI300X's lunch.
It's sim
CUDA (Score:5, Interesting)
Re: CUDA (Score:2)
Can AI write the new tensor kernels using less time and energy than it takes to support a human dev team?
Re: (Score:2)
Yes.
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fsakana.ai%2Fai-cuda-engi... [sakana.ai]
Re: (Score:2)
The reason the MI300X failed to meet expectations isn't because it doesn't use CUDA. It's because the software it does use fucking sucks.
Even if AMD had made a perfectly compatible clean-room reverse engineered CUDA stack for their card, it would probably still suck.
Re: (Score:2)
How much manpower is a large organization going to invest in writing new tensor kernels for some new chip that might have it's API blown away every other dev cycle?
A Chinese organization, in the current geopolitical climate?
Whatever it takes.
What other option do they have?
Re: (Score:2)
Double the VRAM for a similar price and the code is written faster than Huawei can produce the chips. Nobody likes Nvidia dictating the prices. And don't underestimate the huge open source AI community.
Re: (Score:2)
Everyone has an AI accelerator chip. NVidia's secret sauce is CUDA, which has been stable for nearly 20 years.
To oversimplify a bit: you need _one_ person to write matmul for your chip, and you're done. The entire source code for something like DeepSeek starting from raw GPU instructions can probably fit well within 50kb of source code.
It's not a conjecture. Here's one attempt to do that with AMD GPUs, it really does everything starting from command submission: https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fgithub.com%2Ftinygrad%2Fti... [github.com]
Re: (Score:2)
If it's Huawei, and if it is backed by the Chinese government's strategic goals, I'd say probably a lot of companies are going to port their code to that API.
Re: (Score:2)
How much manpower is a large organization going to invest in writing new tensor kernels for some new chip that might have it's API blown away every other dev cycle?
Funny, to me the above describes CUDA quite accurately.
Speaking as someone who has to re-engineer CUDA based AI scripts for every Nvidia card generation just because they cut the support for older CUDA. And there are some that aren't just worth the effort and gray hairs required to make them work...so that Nvidia will just trash them in a year or two in its newest CUDA.
I would love it if what you wrote about CUDA were true. But in the real world, CUDA is not downwards compatible. When a new CUDA version arr
Your tax dollars hard at work! (Score:2)
Good job nVidia (Score:2)
nVidia took the quick money, ballooning their stock price and sales volume by shipping their products off to china and... womp womp... now china reverse engineered enough of it to develop their own. Nice job!
Why didn't the government stop them from doing this! Cried the libertarian shitstains on this website.
Re: (Score:2)
Why didn't the government stop them from doing this! Cried the libertarian shitstains on this website.
I am solidly anti-libertarian (I used to think I was one when I was young, but then I grew up and learned to care about other people) but I also have worked for a chip design firm and know that China has been copying others' designs, re-fabbing them, and selling them for the whole time they could conceivably have been doing that. But on the third hand, how would the government have stopped China from copying our technology? Just not having the chips fabbed in Taiwan wouldn't have done it.
Re: (Score:2)
but I also have worked for a chip design firm and know that China has been copying others' designs, re-fabbing them, and selling them for the whole time they could conceivably have been doing that.
For the stuff that was fabbed there- no doubt. That does however largely preclude just about all advanced processing elements.
But on the third hand
Gripping hand. [catb.org]
Just not having the chips fabbed in Taiwan wouldn't have done it.
Are you asserting that the Chinese are leaking designs from TSMC? I mean it certainly seems plausible- but all evidence is to the contrary.
Re: (Score:2)
Good job. Shouldn't you be on 4chan, or is it still down?
Re: (Score:2)
Hey look, it's one of those shitstains.
Re: (Score:2)
Trying to prevent China... (Score:2)
...from acquiring tech is futile and ultimately self-defeating
There are lots of smart engineers and scientists in China who are really good at overcoming obstacles
Re: (Score:3)
Exactly. Chinese companies were happily building low density semiconductors for home appliances etc while slowly climbing the tech ladder as the next level of tech matured and dropped in price. But then the US imposed the first semiconductor sanctions. The Chinese government took notice, saw where things were headed and within six months had a long term target, a comprehensive strategy and 15 years of guaranteed funding for achieving the target.
Same thing happened in the 1950s when the Soviet Union sus