- Joined
- Aug 19, 2017
- Messages
- 2,595 (0.98/day)
NVIDIA's upcoming GeForce RTX 50-series is expected to make a debut during the CES 2025 show in January. However, NVIDIA serves a lot of markets, and its significant presence is also recorded in the Chinese market as well. Hence, the company has prepared its "Dragon" version of the top-of-the-line GPU versions for Chinese gamers, complying with export regulations that forbid Chinese entities from acquiring GPUs powerful enough to train AI models. NVIDIA appears to be developing "D" variants for both RTX 5090 and RTX 5080 cards, but we only have the confirmation for the GeForce RTX 5090D so far, thanks to MEGAsizeGPU. However, a new leak from the Chinese Chiphell forum points to RTX5090D having the same hardware specification as the regular RTX 5090.
In the last generation, RTX 4090 and RTX 4090D were different, with the RTX 4090D having lower TGP and core count. However, this time around, NVIDIA will physically leave the same expected GB202 "Blackwell" die with CUDA core count and memory, only to lock certain features through firmware. Since the primary goal of export regulations the US is imposing is slowing down Chinese access to GPUs powerful enough to train and inference AI models, we expect to see the RTX 5090D with reduced Tensor core capability and maybe a lowered frequency of the overall chip. This could limit some applications from using these Tensor cores for inference, while NVIDIA's AI features like DLSS could also be a bit slowed down. NVIDIA may find a way to allow DLSS and other gaming-related technologies to operate normally while other general-purpose AI tasks are limited. Readers may recall Low Hash Rate (LHR) cards during the crypto mining boom, which had a similar firmware lock to cap mining hash rate.
View at TechPowerUp Main Site | Source
In the last generation, RTX 4090 and RTX 4090D were different, with the RTX 4090D having lower TGP and core count. However, this time around, NVIDIA will physically leave the same expected GB202 "Blackwell" die with CUDA core count and memory, only to lock certain features through firmware. Since the primary goal of export regulations the US is imposing is slowing down Chinese access to GPUs powerful enough to train and inference AI models, we expect to see the RTX 5090D with reduced Tensor core capability and maybe a lowered frequency of the overall chip. This could limit some applications from using these Tensor cores for inference, while NVIDIA's AI features like DLSS could also be a bit slowed down. NVIDIA may find a way to allow DLSS and other gaming-related technologies to operate normally while other general-purpose AI tasks are limited. Readers may recall Low Hash Rate (LHR) cards during the crypto mining boom, which had a similar firmware lock to cap mining hash rate.
View at TechPowerUp Main Site | Source
