Nvidia Modifies: The RTX 4090D and the Changing Rules Regarding GPU Exports
The GeForce RTX 4090 GPU from Nvidia is once again available for purchase in China, albeit with a less powerful configuration to abide by US export prohibitions to the Middle Kingdom.
The device, known as the RTX 4090D, debuted on Nvidia’s Chinese website on Thursday. Its performance is around 10.94 percent less than that of the model that Nvidia released in late 2022. This is manifested as a reduced number of CUDA cores—14,592 against 16,384 on models distributed outside of China.
Additionally, Nvidia informed The Register today that the 4090D variant’s tensor core count has been reduced by a comparable amount, from 512 to 456. Beyond this, the card remains mostly unaltered, with a massive 384-bit memory bus, 24 GB of GDDR6x memory, and peak clock speeds rated at 2.52 GHz.
As it was previously noted, once the most stringent export rules ever published by the Biden Administration were published in October, the RTX 4090 was the sole consumer graphics card that could no longer be purchased in the United Kingdom.
The card’s performance restrictions for consumer cards with a total processing performance (TPP) of more than 4,800 were just about exceeded, which was the issue. This figure is computed by multiplying by the bit length of the operation and doubling the maximum number of dense tera-operations per second, whether they are integer or floating point.
With a TPP of 5,285 for the original 4090, Nvidia required a permission from the US government in order to market the well-liked gaming card in China. Recall that the performance density criteria, which limits the sale of considerably less powerful datacenter cards like the Nvidia L4, does not apply to consumer cards.
In actuality, reducing performance by 10.94 percent is sufficient to move the card beneath the threshold that calls for the Bureau of Industry and Security (BIS) in the United States to take export license consideration into account.
Nvidia claimed that despite being slower, the performance is still far lower than you may imagine.
“In 4K gaming with ray tracing and deep-learning super sampling (DLSS), the GeForce RTX 4090D is about five percent slower than the GeForce RTX 4090 and it operates like every other GeForce GPU, which can be overclocked by end users,” an Nvidia spokesperson says as per report
Given that the card can be overclocked, users who are prepared to make some minor tweaks may be able to regain some of the performance that was lost due to the lower core count.
Nvidia has already reduced the performance of its graphics cards in order to abide by US export laws. After a series of Chinese regulations on the selling of AI accelerators, the American behemoth reduced the interconnect speed of its well-liked A100 units in late 2022 and reintroduced it as the A800. A new iteration of the H100, dubbed the H800, duly replaced the card. (According to Chinese numerology, eight is the luckiest number; the A800 and H800 were not named by mistake.)
Following the announcement of new export limits in October, Nvidia once more produced chips intended for the Chinese market. Washington had the L2, L20, and H20 limbo dancing directly beneath his performance hats.
US Commerce Secretary Gina Raimondo was swiftly alerted to Nvidia’s actions and cautioned chipmakers not to test those limits. She declared at a forum in early December, “I’m telling you, if you redesign a chip around a particular cut-line that enables them to do AI, I am going to control it the very next day.”
Subsequently, Raimondo told Reuters that Nvidia and the Commerce Department were collaborating closely to make sure that no GPUs or AI accelerators that could jeopardize US national security were supplied to China, but that the chipmaker was allowed to conduct business there. ®