NVIDIA's China-exclusive, RTX 6000D Blackwell GPU, has been tested in Geekbench & reveals a largely cut-down design.
NVIDIA Cuts Down Cores, VRAM & Clocks For Its China-Exclusive RTX 6000D Blackwell GPU
A few months ago, when the US initiated aggressive export controls for AI hardware, being sent out to China, NVIDIA decided to make a brand new SKU to counter those restrictions. It was later confirmed that NVIDIA's new GPU would be called the RTX 6000D Blackwell and will be a different SKU than the one on offer globally.
The NVIDIA RTX PRO 6000 comes in three flavors currently, which include a standard variant, a Max-Q variant, and a server variant. All three feature similar GPU configurations but with different TDP & clock targets. The RTX 6000D for China seems to be a significantly cut-down design, though.
In terms of specifications, the NVIDIA RTX 6000D Blackwell GPU for China features 156 SMs or 19,968 CUDA cores. These are 17% fewer cores versus what is featured on the original model, which offers 188 SMs or 24,064 CUDA cores. The cut-down doesn't end here as the RTX 6000D also features less VRAM, reportedly at 83 GB versus the 96 GB featured on the standard model. We believe that there's an error with the reported VRAM count, as NVIDIA is using 3 GB chips and 84 GB would align perfectly with such VRAM modules. This would indicate a 448-bit bus interface versus the 512-bit bus featured on the full-fat model.
The card also features lower clocks, reported at 2430 MHz, while the standard RTX PRO 6000 configs operate beyond 2600 MHz clock speeds. We don't know the exact TDP for this model, but we do know its performance based on the leak.
NVIDIA's RTX 6000D GPU scored 390,656 points in the OpenCL benchmark test at Geekbench 6, which is lower than the 450-500K score that the full-fat configuration scores. The numbers range from test config to test config, but the GPU was tested on an AMD EPYC system comprising two EPYC 9654 192-core CPUs.
NVIDIA's RTX 6000D Blackwell GPU hasn't seen a good response in the Chinese AI market due to its cut-down specs and value. China has recently doubled down on domestically produced AI chips and is also procuring AI chips through illegal means. Even consumer parts are being outfitted with custom coolers and higher VRAM for AI purposes.
NVIDIA RTX PRO "Blackwell" GPU Lineup:
| Graphics Card | GPU Cores | AI TOPS | FP32 / RT Compute | VRAM | Memory Bus / BW | Form Factor | TDP |
|---|---|---|---|---|---|---|---|
| RTX PRO 6000 | 24064 (GB202) | 4000 | 125 / 380 | 96 GB GDDR7 | 512-bit / 1792 GB/s | Dual-Slot / Extended | 600W |
| RTX PRO 6000 Max-Q | 24064 (GB202) | 3511 | 110 / 333 | 96 GB GDDR7 | 512-bit / 1792 GB/s | Dual-Slot / Full | 300W |
| RTX PRO 6000D | 19968 (GB202) | TBD | TBD | 84 GB GDDR7 | 448-bit / 1568 GB/s | Dual-Slot / Full | TBD |
| RTX PRO 5000 | 14080 (GB202) | 2064 | 65 / 196 | 48-72 GB GDDR7 | 384-bit / 1344 GB/s | Dual-Slot / Full | 300W |
| RTX PRO 4500 | 10496 (GB203) | 1687 | 50 / TBD | 32 GB GDDR7 | 256-bit / 896 GB/s | Dual-Slot / Full | 200W |
| RTX PRO 4000 | 8960 (GB203) | 1178 | 37 / 112 | 24 GB GDDR7 | 192-bit / 672 GB/s | Single-Slot / Full | 140W |
| RTX PRO 4000 SFF | 8960 (GB203) | 770 | 24 / 73 | 24 GB GDDR7 | 192-bit / 432 GB/s | Dual-Slot / Half | 70W |
| RTX PRO 2000 | 4352 (GB206) | 545 | 17 / 52 | 16 GB GDDR7 | 128-bit / 288 GB/s | Dual-Slot | 70W |
News Source: Benchleaks
Follow Wccftech on Google to get more of our news coverage in your feeds.
