Nvidia’s GeForce RTX 4090 GPU is back on sale in China. A less-featured, configured version designed to comply with U.S. restrictions on export to the Middle Kingdom.
The device, called the RTX 4090D, appeared on Nvidia’s China-market website on Thursday and boasts about 10.94 percent lower performance than the model Nvidia announced in late 2022. This manifests itself in a lower core count of 14,592 cores compared to 16,384 CUDA cores in the commercially available version. Outside of China.
Nvidia also said register Now, the card’s tensor core count has also been reduced by a similar margin from 512 to 456 in the 4090D variant. Other than this, the card is largely unchanged, with a peak clock speed of 2.52 GHz, 24 GB of GDDR6x memory, and a fat 384-bit memory bus.
As we reported at the time, the RTX 4090 was the only consumer graphics card banned from sale in the Middle Kingdom following the toughest set of export restrictions announced by the Biden administration in October.
The problem was that this card slightly exceeded the performance limit for consumer cards with a total processing performance (TPP) of over 4,800. This number is calculated by multiplying the maximum number of dense tera operations per second (floating point or integer) by the bit length of the operation.
The original 4090 had a TPP performance of 5,285. This means that Nvidia needed a U.S. government-issued license to sell this popular gaming card in China. Note that consumer cards are not subject to performance density standards that limit the sale of much less powerful data center cards like the Nvidia L4.
Coincidentally, a 10.94% reduction in performance is enough to bring this card to the U.S. Bureau of Industry and Security (BIS) requirements for export license consideration.
Slow Nvidia claimed to have poor performance, but his performance is much lower than you think.
“For 4K gaming with ray tracing and Deep Learning Super Sampling (DLSS), the GeForce RTX 4090D is approximately 5% slower than the GeForce RTX 4090 and performs similarly to all other GeForce GPUs that end users can overclock. ,” an Nvidia spokesperson said in an email.
The card’s ability to be overclocked means that some of the performance lost due to lower core counts could be recovered with a little tinkering on the part of the customer.
This isn’t the first time Nvidia has weakened the performance of its graphics cards to comply with US export regulations. In late 2022, following a series of restrictions on the sale of AI accelerators in China, the American giant reduced the interconnect speed of his popular A100 unit and relaunched it as A800. This card was followed by a new version of his H100, predictably called H800. (8 is the luckiest number in Chinese numerology. The naming of A800 and H800 is no coincidence.)
New export restrictions announced in October forced Nvidia to once again offer chips designed for the Chinese market. L2, L20, and H20 limbo danced directly under Washington’s performance cap.
Nvidia’s antics quickly caught the attention of U.S. Secretary of Commerce Gina Raimondo, who warned chipmakers not to try these bans. “What I’m saying is, if I redesign the chip around a certain cut line that allows me to do AI, the next day I’m going to be controlling it,” she said in December. He spoke at a forum earlier in the year.
Raimondo later told Reuters that the Commerce Department is working closely with Nvidia to ensure that GPUs and AI accelerators that pose a threat to U.S. national security are not sold to China, but that the chip house can do business in China. He said it should be done. ®