GeForce RTX 4090D is equipped with an AD102-250 chip

Yesterday, reports emerged that Nvidia plans to release a China-specific version of its GeForce RTX 4090D graphics card, adapting to the U.S. government’s new export controls on cutting-edge artificial intelligence (AI) chips. This new model, set to replace the restricted GeForce RTX 4090, is expected to arrive early next year.

Subsequent online discussions revealed that the RTX 4090D is equipped with an AD102-250 chip, differing from the AD102-300/AD102-301 chips used in the RTX 4090. While specific specifications remain unclear, it’s certain that the RTX 4090D’s performance, though less than the RTX 4090, will comfortably exceed that of the RTX 4080. Its performance comparison with the upcoming RTX 4080 SUPER, however, is yet to be determined.

Clearly, the RTX 4090D must comply with the Total Processing Power (TPP) limit of 4800. The current RTX 4090’s TPP, whether in FP8 or FP16, stands at 5286, exceeding the limit by about 10%, while the RTX 4080’s TPP is 3117, still a fair distance from the restriction. If Nvidia aims to reduce the TPP from the RTX 4090, assuming a core frequency of 2.7 GHz, it would mean reducing the number of Streaming Multiprocessors (SM) from 128 in the RTX 4090 to at least 108, bringing the TPP down to 4778.

Additionally, the regulations include a metric called Performance Density (PD, TPP divided by chip area). This is why low-performance products like L4 compute cards are restricted, with a TPP of 1936 and a PD of 6.6. If the TPP reaches 1600 and PD hits 5.92, it would also face restrictions. However, this metric does not apply to consumer GPUs; otherwise, the entire GeForce RTX 40 series wouldn’t be sellable in China, as these Ada Lovelace architecture GPUs all have a PD exceeding 6.0.

It is certain that Nvidia will need to appropriately scale down the specifications of the RTX 4090, primarily by reducing the number of SMs, Tensors, and CUDA configurations. However, the RTX 4090D must maintain a certain gap from the RTX 4080 SUPER, even preserving some overclocking space to avoid hitting the restriction. This will be a significant test of Nvidia’s fine-tuning skills.