ChinAI #196: Chinese Reactions to the Nvidia and AMD Chip Ban
Greetings from a world where…
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: The implications of the Nvidia and AMD chip ban on China’s AI ecosystem
Context: Last Wednesday, U.S. officials told Nvidia and Advanced Micro Devices (AMD) that they should stop exporting certain high-end chips to China. Specifically, the ban covered AMD’s M100 and M200 chips, as well as Nvidia’s A100 and H100 chips. This week’s feature translation comes from xinzhiyuan [AI New Era], which provides one window into Chinese reactions to the ban.
Key Takeaways: These moves are seen as directly targeting China’s AI ecosystem.
These chips are especially useful for training large AI models, and the article frames this move as trying to “prevent China from leading the world in the field of artificial intelligence.”
Using Nvidia’s H100s to train a model the size of GPT-3, for instance, would offer a 6.3 times improvement over the A100 (which is currently the state-of-start).
I laughed at this line: “The documents show that the ban targets China, Hong Kong and Russia. However, Nvidia has no paying customers in Russia.” I’m totally on board with implementing a globally consistent approach for technology controls to avoid singling out China — but when it’s so easily seen through, let’s just be up front about it.
Some reactions from Chinese netizens cited by the article and in the comments:
One comment: “Our self-made chips are already on the road, ol' America helped us again.” The top comment in the discussion about the article: “Currently Sugon’s GPU (Sugon calls it a DCU) is about 60% of the performance of A100. It’s just that the (surrounding) ecosystem is not ideal. Even though it’s not that easy to use, it can still do high-performance computing. If the U.S. overthinks it like this, then this will cause China to perfect its own ecosystem. In fact, this is a good thing…Once Sugon’s DCU gets more complete, then Nvidia will be no big deal.”
However, according to one icsmart.cn [芯智讯] analysis cited in the article, there’s still a huge gap between Chinese GPUs and Nvidia/AMD.
Another top comment extrapolates to what could come next: “Maybe the next step will be banning deep learning frameworks torch and tensorflow, this will reveal the importance of Baidu’s self-developed PaddlePaddle framework.”
When it comes to U.S.-China technology competition, the benefits of the “promote” plank will always outweigh the “protect” plank. And, when we reflect on moves like this one thirty years from now, there’s a decent chance that such “protect” actions were counterproductive.
One detail that stood out to me about the notices that AMD and Nvidia received: a directive to collect statistics on previous shipments and customer lists of chips. This is an essential step for end-use verification for export control. My guess is that very few of these chips are going to military applications in any meaningful sense, but if anyone can show me good evidence to the contrary, I’m willing to admit I’m mistaken.
ChinAI Links (Four to Forward)
Should-read: CSET August Research Roundup
Got a chance to catch up on CSET’s August research roundup, which covers a wide range of impressive work. One thing to highlight is a new Country Activity Tracker that includes metrics on countries’ AI ecosystems.
The Reuters coverage on the Nvidia and AMD export bans has been top-notch. In this article, Jane Lanhee Lee and Stephen Nellis explore the possibility that this ban will only accelerate Chinese progress in high-end chips. This quote was especially compelling:
Jack Dongarra, a distinguished professor of computer science who helps lead the Top500 ranking of the fastest supercomputers says he's seen this scenario play out before. "The U.S. embargoed Intel chips from going to specific places in China that are and were developing high performance computers," he said. "The result was that China designed its chips for its supercomputers."
Should-read: The most surveilled place in America
In The Verge, Gaby Del Valle reports on the use of “high-tech capacity” to secure the border:
The lack of legal migration opportunities has been a windfall for two seemingly disparate groups: the smugglers, increasingly tied to cartels, who migrants hire to get them across the border; and the military and technology contractors who are paid billions to develop tools to catch migrants who try to get through the desert. Between 2008 and 2020, ICE and CBP issued more than $55 billion in contracts for detention, “smart border” technologies, and surveillance tools, a report by the Immigrant Defense Project found. The migrant smuggling industry is much more difficult to quantify, but one 2018 United Nations report estimated its worldwide worth as somewhere between $5.7 and $7 billion. The two industries — one licit and one illicit — need each other to function. Both exist because legal migration is nearly impossible.
Should-read: A New ‘Bumper Sticker’ for Space Satellites
For CSIS, Zhanna Malekos Smith has published a piece on space security, detailing possible Russian and Chinese cyber threats to space systems. Some of the details about Chinese researchers’ worries about SpaceX’s Starlink satellite systems are especially intriguing.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Any suggestions or feedback? Let me know at email@example.com or on Twitter at @jjding99