The U.S. government has tightened export restrictions on high-performance semiconductor chips to China, including the Nvidia RTX 4090D, and is urging South Korea to enforce similar curbs, marking a significant escalation in efforts to limit Beijing's access to advanced technology.

Nvidia's RTX 4090D Caught in Tightening U.S. Export Restrictions to China

In a recent report by Tom’s Hardware, the RTX 4090D was explicitly designed to comply with the U.S. China-imposed export bans several months back. The RTX 4090 exceeded the 4,800 Total Processing Power (TPP) limit by 10%, so Nvidia designed the 4090D to come in under that limit. Surprisingly, the new 70 TFLOPS limit is only 5% lower than the RTX 4090D's 73.5 TFLOPS rating.

While this change appears to be unavoidable, people must question its significance. After the launch of the RTX 4090D, the U.S. government warned Nvidia that its tactics would not go unnoticed, and it has now moved to ban Nvidia's China-exclusive GPU. But does a 5% reduction in the GPU speed limit matter, and if so, what happens when Nvidia creates a new GPU that falls below that limit?

The RTX 4090D is a stripped-down version of the RTX 4090, with 14,592 CUDA cores and a 425W TDP. Compared to the previous RTX 4090, the RTX 4090D has 12.8% fewer CUDA cores and a 5.9% lower TDP. All other core specifications remain identical between the two. The only exception is the base clock, which has been increased slightly to 2.28 GHz from 2.23 GHz.

According to other websites that tested the card, the RTX 4090D is about 10% slower than the RTX 4090 in A.I. workloads but only 5% slower in gaming. Ironically, Nvidia never fully "locked" the RTX 4090D, which allowed Chinese gamers and professionals to overclock it to RTX 4090 performance levels.

The RTX 4090D was designed to comply with America's China export restrictions. These laws were enacted to prevent China and non-NATO countries from acquiring excessive computing power, particularly A.I. processing power, for security purposes. These sanctions have been repeatedly changed over the past few years, first targeting data center chips like the Nvidia A100 and Nvidia H100, but later, the RTX 4090 fell victim to the restrictions as it was 10% "too fast."

Shifting Metrics: U.S. Adjusts Performance Benchmarks for GPU Exports to China

The current metric for calculating the maximum allowable performance is TPP or Total Processing Power. The maximum computed for a given bit depth is calculated by multiplying TFLOPS (or TOPS for integer work) by the number of bits. TPP for FP8 work on Tensor cores on the RTX 4090 is 660.8 * 8 = 5,286 (sparsity is not considered).

The new regulations appear to change the defined limit to include "Weighted TeraFLOPS" but fail to explain what that means. The RTX 4090 offers 82.6 TFLOPS of computing, the RTX 4090D reduces that to 73.5 TFLOPS, and the RTX 4080 Super, at 'only' 52.2 TFLOPS, is the next step down for Nvidia's consumer GPUs, according to the language used. It's also worth noting that these repeatedly lowered limits are beginning to creep on AMD's RX 7900 XTX, which has 61.4 computed TFLOPS.

While the government does not explicitly mention the RTX 4090D as the reason for the new restrictions, it is a safe bet that the card will be phased out shortly. Nvidia might keep the 4090D and rebrand it for the Western market... Alternatively, it could release a new GPU that complies with the sanctions again. Let's call it the RTX 4090 DD "Double Dragon" and give it 108 Streaming Multiprocessors, 13,824 CUDA cores, and possibly a 400W TGP — but most importantly, 69.7 TFLOPS of computing. Then it would become compliant again, shift tens of thousands of units to China, and likely face another cut to its allowed export performance.

The new restrictions will reportedly take effect on April 4, 2024. Nvidia has yet to respond to the lower limits, but it will comply with them. But, fundamentally, if 80 TFLOPS was too fast, and 70 TFLOPS is now too fast, the United States must set a hard limit and stick to it — or it will end up 'banning' GPUs that have long been discontinued. It's doubtful that the restrictions are fully effective, as plenty of Chinese customers are still looking for GPUs, and the hardware will likely continue to flow into the country through indirect channels.

U.S. Seeks South Korean Support to Expand Chip Export Restrictions on China

According to Bloomberg, the U.S. is asking South Korea to impose similar restrictions on semiconductor technology exports to China as Washington has already done. This indicates that U.S. President Joe Biden's administration is stepping up efforts to thwart Beijing's chip ambitions.

According to people familiar with the matter, U.S. officials want South Korea to restrict the flow of equipment and technologies used to manufacture high-end logic and memory chips to China.

One of the sources, who asked not to be identified because the discussions are private, includes logic chips that are more advanced than 14 nanometers and DRAM that is more than 18 nanometers.

This would be consistent with the measures the U.S. Department of Commerce announced in 2022.

Last month, U.S. officials held in-depth discussions with South Korean President Yoon Suk-yeol's administration, according to sources.

While the U.S. is attempting to reach an agreement before a G7 summit in mid-June, South Korean officials are debating whether to meet the U.S. request, partly because China remains a crucial trading partner.

Washington's request to Seoul has never been detailed before. This comes from the United States' new push to persuade allies to limit Chinese firms' semiconductor equipment servicing and restrict spare parts and chip exports to China.

According to Bloomberg News (via Taipei Times), the U.S. has pressed its allies, including South Korea and Germany, to tighten restrictions on China's access to their technology. South Korea is a leader in semiconductor production and spare parts for chipmaking equipment.

The timeline may slip. According to sources, officials from South Korea, Japan, and the United States plan to meet in late June to discuss advanced technology and supply chain cooperation.

South Korean officials are concerned about the potential penalties that Beijing's export controls could impose on significant firms such as Samsung Electronics Co. and S.K. Hynix Inc., which continue to operate in China, Seoul's largest trading partner.

South Korea manufactures some of the world's most advanced logic and memory chips in collaboration with Samsung and Hynix. While its chip equipment suppliers are not as well-known as the U.S.'s Applied Materials Inc. or the Netherlands' ASML Holding NV, local gear manufacturers such as Hanmi Semiconductor Co. and Jusung Engineering Co. remain essential to the Asian country's semiconductor ecosystem.

Photo: Microsoft Bing