This doesn't just ban A100s. It bans those and "any future NVIDIA integrated circuit achieving both peak performance and chip-to-chip I/O performance equal to or greater than thresholds that are roughly equivalent to the A100."
It's a ceiling on the performance that can be exported. In an exponentially scaling industry, this is the equivalent of an announcement in advance of a complete ban on competitive products 2-3yrs out.
Given that both political parties have placed export limits on tech to China, it seems this is the new normal.
One way that companies get around ITAR and similar regulations for things like space grade components that have radiation tolerant features is to only rad test up to what the limits are. The components are probably much more resilient than what they are specced for but by only declaring the legal limit, companies can still export their products. The foreign buyers can do their own testing to prove the components can meet their required levels. I've done this before for a commercial chip that was rumored to be radiation tolerant despite nothing from the vendor saying so and found that it well exceeded my expectations.
I wonder if Nvidia can use a similar loophole where the chips are clocked at lower rates for export to meet the law but can be "overclocked" with hardware modifications.
> Under the Export Control Reform Act of 2018 (50 U.S.C. §§ 4801-4852) (ECRA), criminal penalties can include up to 20 years of imprisonment and up to $1 million in fines per violation, or both.
Chinese brand 8fun (bafang) ebike motors have been sold this way for a long time. AFAIK, the spec for their mid drive motors has been well below output potential.
Reminds me of how in the eighties and nineties many MMIC's topped out at 8GHz, on the datasheet.
As 9GHz X-band capable stuff was more restricted dual use stuff.
Yet ERA-2SM for example had more gain on 10GHz than the 8GHz where the official spec ended at. :)
Sounds like the 80s/90s 208hp “gentleman’s agreement” in Japan. Most of those cars put out 320+hp with trivial bolt-on mods but they were all “208hp nudge nudge wink wink”
> where the chips are clocked at lower rates for export
You'd probably have a hard time claiming that was "peak performance" for the chip if it's underclocked with "nudge-nudge-wink-wink" overclocking available.
China would just import the GPUs from black market of countries not facing embargo, Even specific military spec components facing tight scrutiny end up there and other countries facing embargo[1], When Nvidia comes to learn about it they'll just keep quiet until it leaks out.
Not to mention, H100 is itself made in China and it seems Nvidia is allowed to keep building it there[2].
This could work for GPUs: Instead of selling a 20 TFLOP card for $500 you could sell 2 for $250 each consuming half the power. Wonder how long you could keep this up with exponential scaling though...
It's not just the US government, it's many civilized governments doing these sort of things. And in this particular case this european is grateful for it because it is not about setting ceilings or hindering industry, it's about preventing the abuse of dual use goods. In other words; Not being fsckd by our own stuff.
I worked with a company that upgraded a mainframe cluster not located in the USA. The upgrade topped out a collective rating of that user's installed computing power allowance given for civilian purposes. So, to do the upgrade, US Export controls required the client to decommission some older nodes and that had to be witnessed by US Consular representatives walked through the whole process taking video and photographic evidence. And the agreed method of decommissioning was to haul these old nodes out in to a deserted area and run bulldozers over them in a pit. Not joking. Wish I kept some of the photos though.
Truly a surprise change. The new requirements do not appear to be published in the Federal Register, which is a massive screwup that potentially renders the requirements illegal and invalid. My contacts in the trade compliance world are aghast.
Will be interesting to see if Nvidia challenges it.
I can’t see how this would be effectively enforced - what’s to prevent them from spinning up an arm’s-length subsidiary (i.e. controlled through trustworthy middlemen) outside of US jurisdiction and using that to sell to prohibited markets? Like, exactly what, for instance, crocs are doing with Russia - or Tetley’s did with Teapigs (for different reasons).
Nvidia does their main chip fab in Taiwan AFAIK, along with a large number of other companies. It's one of the reasons tensions over Taiwan are worrying. China could blockade or invade Taiwan and start a "shortage" far worse than anything we've seen recently if they decided it was necessary/worthwhile.
Yeah I mean, great for running inference (and obviously training smaller models). The problem is that Scaling Laws for Transformers show us that scaling model parameter count _up_, is a sure way to improve the model's performance over your task.
So when you're playing nation-state hard-ball, since the architectures are _somewhat_ trivially copied upon being published, that means the best checkpoint goes to the one with the fastest GPU's. If China wasn't behind in deep learning, they will be.
last I checked that's not what's at issue. they can't produce equivalent chips themselves and rely on them for various physics simulations that back weapons research
Eric Schmidt, the former Google chairman, told Reuters in a recent interview that high-end processors should have kill-switches.
“Knowing where the chips go is probably a very good thing. You could for example, on every chip put in essentially a public private key pair, which authenticates it and allows it to work”.
What he won’t tell is that this is already a reality, as I learned after having my air-gapped system and Pixel phone wiped remotely for researching “silent speech interfaces”. There is no security when silicon trojans are in all devices.
Here is how Biren says it stacks up on various machine learning workloads, pitting the BR100 against the Nvidia A100:
…
We presume this is for AI training workloads, not inference. The average speedup over the A100 is around 2.6X. It is not clear if the Nvidia machines were using sparse matrix features, which doubles the throughput, or not. Our guess is they were not.
I'm optimistic that this diversity will unseat the CUDA framework's dominance in high-performance compute. OpenCL, AMD ROCm, and now "Birensupa" will compete with CUDA; hopefully leading to reduced dependence on any individual manufacturer's proprietary APIs.
Does anyone actually use ROCm? It seems barely supported by AMD, with minimal development, and difficult to run in nearly any scenario. It's been a massive disappointment for someone who generally prefers AMD gpus.
I am not sure what exactly the thought process is here? So NVIDA is losing Chinese market, probably for the foreseeable future. Chinese are one step away from figuring out this technology anyway, if not already. They will not trust us again. It makes them less dependent on us. Oh, and we still need their manufacturing because of the labor costs, environmental issues, regulations and now, energy costs. How would this work exactly? I know it has been done during the Cold War I, but the world looks very different today.
For the same reason why you don't send good money after bad. A line is drawn in the sand and while this isn't significant at first - it compounds over time, and allows businesses to divest, avoid new relationships and allow existing ones to become redundant.
From war-mongering, using arbitrary embargoes as political retaliation, to farcical business practices (e.g. the ARM China saga), to a lop-sided approach to intellectual property and using the state to remove competition, both China and Russia have proven to act in bad faith, not respect rules-based order and not be a good business/manufacturing partner.
At some point one needs to stop feeding the trolls.
>From war-mongering, using arbitrary embargoes as political retaliation, to farcical business practices (e.g. the ARM China saga), to a lop-sided approach to intellectual property and using the state to remove competition
This description fits USA much better than China. Especially the war mongering - China hasn’t invaded anyone for five decades now.
>Chinese are one step away from figuring out this technology anyway
Even if they do they are far away from fabrication.
>They will not trust us again.
China already massively distrusts the West.
America tried the Japanese economic model with China (allow your factories to be exported there, remove trade barriers) and it didn't get them very far.
It's kind of astounding that the Senate can literally legalize insider trading for themselves and nothing comes of it. There's no one pointing out how that's just garden-variety corruption common in every authoritarian, extractive state, but one step removed. A great example of how even in a democratic society, the people cannot realistically expect lawmakers to pass laws against those lawmakers' own economic interests.
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]!
Very worrying to see the restrictions against China escalating;
Especially since China is incredibly advanced in hardware manufacturing and is not at war with the West.
At best, this is just pushing for escalation so they can restrict exports on their side too.
Even, for them. In one of their speech (in France, on a YouTube channel called Thinkerview for those who know), one of their representative quoted a Chinese proverb saying: "Nothing is more favourable than stability, nothing is more harmful than chaos".
Social stability, free markets and long-term peace are in the interest of both parties.
This is what the West tried to do with Russia. Germany especially was heavy on economic ties. "Surely if we're economically tied together, Russia won't risk their prosperity by full-on invading another country!" How well did that turn out for them?
you'd hope people who make these decisions also include some mid and long term thinking. On that time scale any politicization of trade is going to accelerate competitors and incentivize everyone to look towards alternatives. That's basically what's happened in software already.
A major reason the US has had sustained influence in the world without using much force is because it has not weaponized trade. It's the reason for the strength of the dollar, the dominance of US tech and finance, and so on. Throwing that away to merely stagger a country that's destined to catch up eventually doesn't seem wise.
I'm not a foreign policy expert, but the PRC seems to have very aggressive territorial ambitions. Its nine dash line incorporates vast swathes of land and water that international law recognizes as belonging to other countries.
To strengthen its claims in the West Pacific, the PRC has embarked on an unprecedented island building program in the region, and turned those islands into military outposts.
Here's a good breakdown of the history of PRC claims in the West Philippine Sea:
There’s a long history of the top tier of computing hardware being export controlled. This is more of a recognition (a bit late eh?) of the utility of these graphics cards, though really it’s about time we started calling them something else.
Since we have Cold War II (with the hot war still not fought by NATO), we should expect similarities to Cold War I: bans on high-tech exports, classifying more stuff, even public lies in an attempt to misdirect the adversary (see [SDI]). Good thing if encryption is not declared munition as it was back then [W].
An AI arms race will be limited by the silicon it's built on. The arms race will be in the fabrication of the chips. You can only scale old nodes so far.
The cheapest solution out of all the possibilities is to just use a back channel to get them, of course.
It would take them years to acquire the GPUs at the scale that they need them. By the time they do that they're 2 generations behind. It's about adding a barrier.
This doesn't sound right to me. As a back-of-the-envelop calculation how much realistically would say the Russian government need? Even if they need in the league of 10,000s annually I think it would be easy for them to source this from a handful of friendly countries in a couple of months.
For example they could set-up a couple of "mining startups" in Dubai, Egypt, Serbia or India. Then buy 1000s of GPUs under the companies name and then put them in a yacht and ship to Moscow.
I literally know people who fly into Russia with entire disassembled gaming rigs in their suitcases (and do their best to avoid customs) from Turkey, etc. I'm sure if the Russian government needed these, they'd set up an operation to let them through. Russia is not as closed off as people seem to imagine.
Depends on what you're really trying to do. Having to pay that putative 30% markup on hardware is going to be a crippling disadvantage for a China-based AI product, for example.
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.
And then Kazakhstan buys them from the Saudis or from Turkey, who are de facto US allies (Turkey is even a NATO member, a big one). The US can try and do something about it, as supposedly the CIA has just tried to do a couple of weeks ago in connection to the recent sanctions against Russia, but what do you when the Turks tell you to shove it? (which is what Turkey's response to the CIA demands was). This is one of the "best" sources I could find in English [1] because, obviously, the Western media has not reported on it.
any vendor has an owner who very likely enjoys traveling to the west, banking there etc. risking being on some black list for a marginal gain is not very attractive prospect
Readit News
It's a ceiling on the performance that can be exported. In an exponentially scaling industry, this is the equivalent of an announcement in advance of a complete ban on competitive products 2-3yrs out.
Given that both political parties have placed export limits on tech to China, it seems this is the new normal.
I wonder if Nvidia can use a similar loophole where the chips are clocked at lower rates for export to meet the law but can be "overclocked" with hardware modifications.
I work in CPU design and I wouldn't risk it.
Yet ERA-2SM for example had more gain on 10GHz than the 8GHz where the official spec ended at. :)
You'd probably have a hard time claiming that was "peak performance" for the chip if it's underclocked with "nudge-nudge-wink-wink" overclocking available.
Not to mention, H100 is itself made in China and it seems Nvidia is allowed to keep building it there[2].
[1] https://www.reuters.com/technology/chip-challenge-keeping-we...
[2] https://www.cnbc.com/2022/09/01/nvidia-says-us-government-al...
This is not really new but NVidia was caught by a surprise change in the limits.
Other examples can be found here: https://www.wassenaar.org/app/uploads/2021/12/Public-Docs-Vo...
Will be interesting to see if Nvidia challenges it.
Worked on a three 50U rack recently with quad T4's per 2U and we're not even through 10% of that order yet.
So when you're playing nation-state hard-ball, since the architectures are _somewhat_ trivially copied upon being published, that means the best checkpoint goes to the one with the fastest GPU's. If China wasn't behind in deep learning, they will be.
Edit: why disagree?
“Knowing where the chips go is probably a very good thing. You could for example, on every chip put in essentially a public private key pair, which authenticates it and allows it to work”.
hxxps://www.reuters.com/technology/chip-challenge-keeping-western-semiconductors-out-russian-weapons-2022-04-01/
What he won’t tell is that this is already a reality, as I learned after having my air-gapped system and Pixel phone wiped remotely for researching “silent speech interfaces”. There is no security when silicon trojans are in all devices.
Fabbed at TSMC. (I guess we are about to see if SMIC can do their 7nm at scale.)
https://www.nextplatform.com/2022/08/25/china-launches-the-i...
CHINA LAUNCHES THE INEVITABLE INDIGENOUS GPU
…
Here is how Biren says it stacks up on various machine learning workloads, pitting the BR100 against the Nvidia A100:
…
We presume this is for AI training workloads, not inference. The average speedup over the A100 is around 2.6X. It is not clear if the Nvidia machines were using sparse matrix features, which doubles the throughput, or not. Our guess is they were not.
https://www.servethehome.com/biren-br100-gpu-for-datacenter-...
From war-mongering, using arbitrary embargoes as political retaliation, to farcical business practices (e.g. the ARM China saga), to a lop-sided approach to intellectual property and using the state to remove competition, both China and Russia have proven to act in bad faith, not respect rules-based order and not be a good business/manufacturing partner.
At some point one needs to stop feeding the trolls.
This description fits USA much better than China. Especially the war mongering - China hasn’t invaded anyone for five decades now.
Even if they do they are far away from fabrication.
>They will not trust us again.
China already massively distrusts the West. America tried the Japanese economic model with China (allow your factories to be exported there, remove trade barriers) and it didn't get them very far.
Where do you think most of the electronics is manufactured?
Then they can ban their products from the US
... or we're governed by morons. Take your pick.
Nice that this comes a month after the Pelosis cleared their NVDA holdings. Just a coincidence, I'm sure.
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
One the applications they are hungry for.
Especially since China is incredibly advanced in hardware manufacturing and is not at war with the West.
At best, this is just pushing for escalation so they can restrict exports on their side too.
Even, for them. In one of their speech (in France, on a YouTube channel called Thinkerview for those who know), one of their representative quoted a Chinese proverb saying: "Nothing is more favourable than stability, nothing is more harmful than chaos".
Social stability, free markets and long-term peace are in the interest of both parties.
A major reason the US has had sustained influence in the world without using much force is because it has not weaponized trade. It's the reason for the strength of the dollar, the dominance of US tech and finance, and so on. Throwing that away to merely stagger a country that's destined to catch up eventually doesn't seem wise.
What I am calling for: closer ties and negotiations and collaborative work.
Dead Comment
To strengthen its claims in the West Pacific, the PRC has embarked on an unprecedented island building program in the region, and turned those islands into military outposts.
Here's a good breakdown of the history of PRC claims in the West Philippine Sea:
https://youtu.be/H3NzRvvnjRQ
If the image generation market is getting too saturated, maybe OpenAI can sell something to the government. No one has deeper pockets.
[SDI]: https://en.wikipedia.org/wiki/Strategic_Defense_Initiative
[W]: https://en.wikipedia.org/wiki/Export_of_cryptography_from_th...
The cheapest solution out of all the possibilities is to just use a back channel to get them, of course.
I doubt, though, that they will restrict the export of the raw material for creating AI: data.
Hard to see any particular country having the desire or where for all to undertake such an endeavour.
The only way to really do this is to geo-fence disable the chips.
For example they could set-up a couple of "mining startups" in Dubai, Egypt, Serbia or India. Then buy 1000s of GPUs under the companies name and then put them in a yacht and ship to Moscow.
Replacing CUDA programs will take work, but they have enough people for that and this is an easier task than designing a GPU.
Just buy less GPUs and run them for more time. It's not like GPU clusters are running at 100% anyway.
(Also, implying China does not have enough money to buy them like that)
[1] https://www.telesurenglish.net/news/CIA-Allegedly-Targets-Tu...
So these chips which are banned from going from the US to China, do already go from China to the US.
https://www.csis.org/analysis/what-us-one-china-policy-and-w...