As promised, the US gov has whacked another GPU mole headed for China with new limits cooked up to kill the RTX 4090D
The US government has revised its sanctions (via Videocardz) on the level of GPU power it will allow to be exported to China. And it seems aimed specifically at stopping Nvidia selling its China-exclusive version of the RTX 4090 graphics card. With its “weighted TFLOPS” limit now being dropped down to 70 TFLOPS, that means both Nvidia’s RTX 4090D and its H20 data centre GPUs will be unable to be exported to China without a specific license.
Which I expect is probably pretty hard to come by.
This game of whack-a-mole has been going on since the US government introduced its initial sanction list back in October of last year and a host of data centre GPUs were blocked, as well as the mighty RTX 4090.
At the time the commerce secretary, Gina Raimondo said: “The updates are specifically designed to control access to computing power, which will significantly slow the PRC’s (People’s Republic of China) development of next-generation frontier model, and could be leveraged in ways that threaten the U.S. and our allies, especially because they could be used for military uses and modernization.”
Basically, the US doesn’t want serious AI inferencing power in the hands of the PRC. But, since those sanctions specify a Total Power Performance (TPP) limit for GPUs allowed to be shipped to China, and it’s a rather large market for Nvidia to sell graphics silicon into, the green team set to work creating a Dragon to get around those sanctions.
The RTX 4090D, or ‘Dragon’, was designed to sneak just under the TPP limits initially set out with some back-of-a-napkin maths designed by the US gov to put a hard limit on what could be sold. Interestingly, it also remained unlocked, meaning that with a little overclocking nous you could still make it perform at the same level as the standard RTX 4090, making the sanctions rather laughable.
That redesign wasn’t a popular move, and ol’ Gina Raimondo especially wasn’t happy. During a National Defense Forum she is quoted as saying: “We cannot let China get these chips, period. We’re going to deny them our most cutting edge technology… If you redesign a chip around a particular cut line that enables them to do AI, I’m going to control it the very next day.”
So, while it’s not necessarily the next day, the US government has now set out to control that RTX 4090D which was redesigned to get around that “cut line.” The new line has been set at 70 “weighted TFLOPS” according to the newly revised sanctions set to come into force on Thursday, 4th April.
Theoretically the RTX 4090D comes in at 73.5 TFLOPS, putting it over this new limit and suddenly within the bounds of sanctions.
So, what’s to stop Nvidia from redesigning it again, given that the limit is only a paltry handful of TFLOPS below what the RTX 4090D is rated at? Well, nothing other than time and money in terms of reconfiguring the cards in production. Though, there’s also nothing to stop the US from dropping the arbitrary limit once more if Nvidia does go down that road again.
By putting the limit so tantalisingly close, it almost feels as though the US government is baiting Nvidia to do it again. So, it will be interesting to see if Jen-Hsun, the Taylor Swift of Tech, decides the China AI market is still lucrative enough to design another chip to drop below that latest TPP limit.
Though, as Tom’s Hardware notes, if this game continues on, and those TPP numbers keep on dropping then even AMD hardware could get bitten by such sanctions, with its RX 7900 XTX coming within reach at 61.4 TFLOPS of processing grunt.
Won’t somebody think of the GPUs?