NVIDIA's China-exclusive RTX 6000D Blackwell GPU has been tested, revealing a significantly cut-down design compared to its global counterparts. This new SKU was introduced in response to US export controls on AI hardware bound for China. The RTX 6000D features 17% fewer cores, 14% less VRAM, and lower clock speeds, making it a more modest version of the RTX PRO 6000. The reduced specifications have not been well-received in the Chinese AI market, as the country is increasingly focusing on domestically produced AI chips and even modifying consumer parts for AI applications.
The RTX 6000D's specifications include 156 SMs or 19,968 CUDA cores, a significant reduction from the 188 SMs or 24,064 cores in the standard model. The VRAM is also reduced to 83 GB, which might be an error, as NVIDIA typically uses 3 GB chips, making 84 GB a more likely figure. This would result in a 448-bit bus interface, a decrease from the 512-bit bus in the full-fat model. The card's clocks are reported at 2430 MHz, lower than the standard RTX PRO 6000's speeds.
In terms of performance, the RTX 6000D scored 390,656 points in the OpenCL benchmark test, which is lower than the full-fat configuration's score. The GPU was tested on an AMD EPYC system with two EPYC 9654 192-core CPUs. Despite the cut-down design, the RTX 6000D's performance is still lower than expected, possibly due to the reduced specifications.
NVIDIA's RTX PRO 6000 series offers three variants: standard, Max-Q, and server. While the RTX 6000D is a cut-down version, it remains a distinct SKU from the global RTX PRO 6000 lineup. The reduced specifications and value proposition of the RTX 6000D have not resonated well with Chinese AI market customers, who are increasingly favoring domestically produced AI chips and even modifying consumer parts for AI tasks.
