A teardown video has revealed the PCB design and cooling system of the Nvidia RTX 6000D professional GPU, showing four fewer memory chips and a narrower memory bus. For this China-exclusive model, Nvidia has cut a portion of the memory pool and GPU CUDA cores to fit within the performance limit authorised for exportation.
As its name implies, the RTX 6000D is a special version of the RTX Pro 6000 Blackwell graphics card available on western markets, cut down to comply with the USA’s regulations regarding tech exports to China. To respect these rules, this D-suffixed model’s VRAM capacity is reduced by 12.5%, from 96GB to 84GB, and its CUDA core count by 17%, from 24,064 to 19,968.

As a result, the floating-point performance drops from 126 TFLOPS to 97.04 TFLOPS, while the memory bandwidth shrinks from 1,790GB/s to 1,570GB/s. While the former is the reflection of the lost CUDA cores, the latter is simply a byproduct of the narrower memory bus. As a reminder, the regular RTX Pro 6000 features a 512-bit bus, whereas the RTX 6000D makes do with a 448-bit bus.

Aside from some slight differences in the power delivery system, the RTX 6000D seems identical to the regular model. Only 28 out of the 32 memory locations are populated with 3GB GDDR7 chips, hence the commensurately narrower memory bus. The memories are spread equally between the front and back in clamshell mode, where two chips share the same memory channel.

While this isn’t the first time we hear about this GPU, as it was previously spotted on a Geekbench test, today’s disassembly video shared by GINNSOD on BiliBili shows the internal layout of the card. For instance, we can see that the RTX 6000D is equipped with the usual two-slot passive heatsink generally found on server-oriented models. This design relies on the server’s built-in fans to push air through its fins to cool the GPU, memory, and VRAMs.
To allow the card to run on a workstation system, the video host swapped this cooler for a liquid-cooling block, which will come in handy to maintain the GPU at a comfortable temperature. Furthermore, the 6000D seems to be limited to around 400W, unlike the regular Pro 6000’s 600W.
For reference, Nvidia has three variants of the RTX Pro 6000: the Server Edition, Workstation Edition, and Max‑Q Workstation Edition, which respectively target 600W, 600W, and 300W. The RTX 6000D doesn’t fit any of these, which indicates that it is a power-limited version of the Server Edition.
The Nvidia RTX 6000D has been spotted as part of dual-GPU workstation prebuilds for around $26,000, coupled with an AMD Threadripper Pro CPU.

