You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Nvidia said Thursday that the U.S. government told it that it can continue to develop its H100 artificial intelligence chip in China.
www.cnbc.com
Nvidia on Thursday said the U.S. government will allow it to continue developing its H100 artificial intelligence chip in China. It’s a win for the company after it warned Wednesday that new export restrictions could hamper its operations in the country.
Nvidia said Thursday that the U.S. government told it that it can continue to develop its H100 artificial intelligence chip in China.
www.cnbc.com
Nvidia on Thursday said the U.S. government will allow it to continue developing its H100 artificial intelligence chip in China. It’s a win for the company after it warned Wednesday that new export restrictions could hamper its operations in the country.
Nvidia said Thursday that the U.S. government told it that it can continue to develop its H100 artificial intelligence chip in China.
www.cnbc.com
Nvidia on Thursday said the U.S. government will allow it to continue developing its H100 artificial intelligence chip in China. It’s a win for the company after it warned Wednesday that new export restrictions could hamper its operations in the country.
Is such a ban really enforcable other than making life more difficult for the sanctioned? I mean, this is stuff you can buy from retail like newegg. Besides, the Nvidia gaming cards are also AI capable iirc.
My understanding is that Nvidia does not have "key" GPU R&D in China, whatever that means. It's mostly the sort of R&D you do in cooperation with customers, of which Nvidia has many in China, and I suspect that much of their business is now threatened by regulatory concerns. Here's an article quoting Bill Dally, Nvidia's chief scientist, from 2020 about their intentions to grow an R&D center.
Nvidia recently held its China annual CPU technology conference online, and its chief scientist Bill Dally introduced how Nvidia's AI-accelerated ... - Nvidia, R&D
www.chinatechnews.com
I think Nvidia's major international R&D growth site is in Israel.
Chip giant's R&D activities in Israel are already the firm’s largest outside the US; company currently employs about 3,000 people across 7 centers nationwide
Bear Moo Moo. I don't think we are talking newegg, games, cards. I think we are talking embedded GPUs, vector extensions, and routes to HBM3. Missiles like that stuff.
US allows AI R&D in China... unbelievable. Does Marco know that?
Is such a ban really enforcable other than making life more difficult for the sanctioned? I mean, this is stuff you can buy from retail like newegg. Besides, the Nvidia gaming cards are also AI capable iirc.
These bans are on data center GPUs which are sold direct by the manufacturers. As for replacing them with gaming cards, you could. However they would be WAY less efficient and require more servers to get the same computational power.
I continue to think that all this will do is drive China to create its own CPU and GPU designs that meets its needs, and these designs will probably compete in the world market with the US companies. For a few years, perhaps several, the Chinese products probably won't be very competitive, but if multiple US companies can create datacenter GPU designs, so can Chinese companies. IMO, this will be a chip design case of what doesn't kill you makes you stronger.
I continue to think that all this will do is drive China to create its own CPU and GPU designs that meets its needs, and these designs will probably compete in the world market with the US companies. For a few years, perhaps several, the Chinese products probably won't be very competitive, but if multiple US companies can create datacenter GPU designs, so can Chinese companies. IMO, this will be a chip design case of what doesn't kill you makes you stronger.
They are already trying, and in some endeavors they do reasonably well and in others they are woefully terrible. Just like with fab work chip design requires a whole village of well coordinated engineers and all of their tribal knowledge to succeed. Poaching a few engineers won't build that level of competency on the time frames that the PRC would like. A bigger concern is the actual processes this chips will be printed on. It doesn't matter if they have the best designers in the world if all they have is an unlicensed copy of n16 with worse yields and low volumes. Their n7 copy will probably eventually have passable yields but until they can produce their protype litho tools at a large scale (and those tools don't have like half the throughput and 1/3 the reticle size), then these nodes will only ever exist on paper. If these challenges are overcome then where do they go from there; hextuple patterning for a 5nm class node? Because I can guarantee you that we won't see production ready EUV (or any EUV for that matter) at a large scale for years.
They are already trying, and in some endeavors they do reasonably well and in others they are woefully terrible. Just like with fab work chip design requires a whole village of well coordinated engineers and all of their tribal knowledge to succeed. Poaching a few engineers won't build that level of competency on the time frames that the PRC would like. A bigger concern is the actual processes this chips will be printed on. It doesn't matter if they have the best designers in the world if all they have is an unlicensed copy of n16 with worse yields and low volumes. Their n7 copy will probably eventually have passable yields but until they can produce their protype litho tools at a large scale (and those tools don't have like half the throughput and 1/3 the reticle size), then these nodes will only ever exist on paper. If these challenges are overcome then where do they go from there; hextuple patterning for a 5nm class node? Because I can guarantee you that we won't see production ready EUV (or any EUV for that matter) at a large scale for years.
I hope you're correct, and I hope I'm giving China too much credit in process development. Their N7 work perhaps impressed more than it should. A subsidized N7 that yielded "enough", regardless of cost, could tide them over for several years, perhaps long enough to get DUV tools working. I'm sure you'll agree, in technology races it takes less energy to follow than lead, because you get to learn from the R&D successes and failures of your competitors, even just issued patents. And the more we try to suppress them the easier it is to whip up patriotism, even in the ranks of academia and engineers.
www.cnbc.com
www.chinatechnews.com
www.timesofisrael.com
