Live data from Hacker News

US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

sec.gov

321–330 of 645 posts

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#321
post #284

Earlier quoted context omitted.

It's also going to be a 30% additional discount on Chinese TPU equivalent chips, especially since they have 7nm production now. I foresee it backfiring.

That depends on how SMIC's ramp goes. Right now they're shipping what amount to samples only. And even then there's no chips to fab absent more development work. Making a ML accelerator is hard . If AMD and Intel (actually Google and Tesla seem closer) can't beat NVIDIA, you really think Goodix is going to get it done? High tech is still high tech for a reason. Sure, in the long term China is going to get there event…

Are they? Do you know why China stopped submitting any supercomputers to Top500 despite widespread reports of two exascale supercomputers coming online? And why would they prioritize commercial chip sales over national security projects?

AMD and Google have already beat Nvidia at the ML accelerator hardware game in the past. Google TPUs have had the edge over Nvidia cards for a few years, and AMD spent almost a decade with an edge in compute over Nvidia at every level. It's just that Google TPUs aren't commercially available and AMD GPUs lacked the software stack and critical mass.

Actually building a fast ML accelerator is far from a monumental task. It's difficult, but well within the reach of a country which fully designed a Top10 processor by itself. And now has the means to manufacture super high end chips.

It's not a case of intermediate decade or two. The Chinese are pretty much already there. They already have at least one indigenous exascale supercomputer and their own 7nm process. They can make an ML accelerator comparable to the A100 if they want to. Believe it or not, there are only so many ways of doing matrix multiplication in hardware. The hard part is lithography, and they did it.

EDIT: I forgot to mention, but SMIC wanted to keep their 7nm production secret to reduce incentives for the US to force ASML to stop 28nm lithography machines, keeping the impression that banning the sale of EUV machinery was enough to stop their advances. That's why they mix everything under 28nm in their financial reports, to show their very strong profitability without tipping their hand.

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#322
post #297

Earlier quoted context omitted.

You say that as if it didn't just turn out that China has been producing 7nm chips for at least a year.

I think it's generally accepted that this was a copy paste of the TSMC process. I think we'll have to wait to see if they can advance this, on their own. As the other person pointed out, a good indicator would be yield.

It's a copy past of the TSMC process because they were able to offer far higher higher salaries and attract TSMC talent. Given the reality that they have a lot of that talent and more each year and given the excellent track record of SMIC leadership, they will probably be able to advance on this, limited by the lithography machines they can procure.

If their yield was abysmal, no one would be willing to buy the chips since they aren't selling them at a loss and have plenty of public sector demand. The fact that they were able to sell them commercially and thus competitively means that their yields are okay at least.

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#323

Earlier quoted context omitted.

How long would it take one A100 to crack RSA 2048?

Even if you had your GPU-packed Kilo-Google per person multiplanetary Gigagalactic computer guessing numbers for 37 times the age of the universe, it would still have a 1 in 32 billion chance of cracking. https://youtu.be/S9JGmA5_unY?t=32

Your video is talking about 256-bit symmetric encryption. RSA is asymmetric encryption. RSA 2048 is ~112 bits of protection when comparing to symmetric keys, kinda. We're still talking a long time with current computing hardware though.

> For example, if SHA-256 is used with RSA and a 2048-bit key, the combination can provide no more than 112 bits of security because a 2048-bit RSA key cannot provide more than 112 bits of security strength - NIST SP 800-57 PART 1, Page 58

Warning, PDF: https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.S...

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#324
post #317
post #315

I'm confused as to how the ban of exports to China is justifiable. Is this ban not some form of aggression? I can't see how the ban will achieve it's aims of "limiting" China nor do I see why it's acceptable for the US to do this.

I am also confused. I understand in times of war or conflict, but this seems like deliberate economic sabotage.

That’s exactly what it is. Chips is one of the few advantages the US has left over China. It would be foolish to help them, we are in the early stages of the second cold war.

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#325

Earlier quoted context omitted.

Counting the track records … Vietnam, Afghan … apart from wasting lots of tax payer money and making the people in the area miserable, what else did we achieve?

Those are your go to for examples of the US lacking determination and resolve to stand by allies? Sure the US abandoned the South Vietnamese and the Afghan Governments, but only after two decades of military conflict in support of each. A dragged out conflict over Taiwan hurts the CCP the most.

Please remember the CCP has no feeling … hard to make it feel hurt. However, a dragged out conflict would hurt the people at TW and waste our money … let’s focus on building this country strong and take care the ppl here …

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#328
post #251

Earlier quoted context omitted.

The US government has been setting the ceiling on performance that can be exported since the beginning of the computer industry. This is not really new but NVidia was caught by a surprise change in the limits.

When was this done previously?

There was also a ban export of strong cryptography. So Java couldn’t use string ciphers and algorithms without adding some additional jar which basically just turned the `enableStrongEncryption` flag on. This isn’t necessary anymore but I don’t know what changed legally.

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#329
post #284

Earlier quoted context omitted.

That depends on how SMIC's ramp goes. Right now they're shipping what amount to samples only. And even then there's no chips to fab absent more development work. Making a ML accelerator is hard . If AMD and Intel (actually Google and Tesla seem closer) can't beat NVIDIA, you really think Goodix is going to get it done? High tech is still high tech for a reason. Sure, in the long term China is going to get there event…

Are they? Do you know why China stopped submitting any supercomputers to Top500 despite widespread reports of two exascale supercomputers coming online? And why would they prioritize commercial chip sales over national security projects? AMD and Google have already beat Nvidia at the ML accelerator hardware game in the past. Google TPUs have had the edge over Nvidia cards for a few years, and AMD spent almost a decad…

Your contention is that SMIC is successfully operating a high volume 7nm process... in secret? What potential incentive would they have to do that? This isn't a spy novel; if a company has a money factory they tell their investors about it!

Re: US Government Bans Export of Nvidia A100 and H100 GPUs to China and Russia

#330
Although this seems to only affect the Chinese and Russian governments (as opposed to researchers), I'm generally against economic brinkmanship of this variety. It just festers resentment, and doesn't seem to achieve much: of what little I could find and scan in an hour, it appears that Nvidia GPUs strong-scale pretty well for neural networks ([1], [2]). Although this depends heavily on the NN architecture and training algorithm, Ampere architectures _seem_ to suffer from similar levels of atomic contention as Volta architectures [3]. That said, the throughput for half-precision floating point and integral data is much higher on Ampere [3]. Hopefully their final report includes NVLink bandwidth on Ampere.

So what's stopping an adversary from just buying multiple V100s? Or potentially faster accelerators from countries outside the US [4]?

If you're familiar with GPU strong-scaling, please chime in [5]!

[1] - https://arxiv.org/abs/1807.09161

[2] - https://arxiv.org/abs/2111.04949

[3] - https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s...

[4] - https://arxiv.org/abs/1912.03413

[5] - https://meta.wikimedia.org/wiki/Cunningham%27s_Law

Post reply on HN