RTX PRO 6000 Max-Q underwhelms

This is less a clean “bad GPU” story than a useful warning about buying expensive pro AI hardware before the software stack and benchmark expectations are nailed down.

• –Max-Q keeps the same 96GB GDDR7 headline appeal, but its 300W power cap makes comparisons against the 600W RTX PRO 6000 Workstation misleading
• –LLM inference results depend heavily on llama.cpp build flags, CUDA/toolkit versions, quantization format, backend maturity, and whether Blackwell paths are optimized
• –The Windows and Ubuntu mismatch narrows the likely causes, but cross-platform slowness still does not prove defective hardware without controlled power, clocks, thermals, and known-good benchmark runs
• –Community comparisons against dual 5090s and other RTX PRO 6000 setups are valuable, but apples-to-apples model, quant, context, batch, and backend settings matter more than GPU nameplate alone