NVIDIA’s workstation Blackwell card: 96GB GDDR7 ECC at 1,792 GB/s, 2025,
$8,565 MSRP - and it shares the RTX 5090’s bandwidth number while
carrying three times the memory. Two variants on this site: the 600W
standard card and the 300W Max-Q (same memory, same bandwidth, lower
clocks and price of admission to your PSU).
What it does well:
-
96GB single-slot capacity: 70B dense at 8-bit, 235B-class MoE at
4-bit, in one card with full CUDA - the largest-memory single GPU on
this site that runs the standard consumer runtime stack.
-
Server-grade memory: ECC GDDR7 and datacenter driver support, in a
dual-slot card a workstation can hold.
Where it falls short: $8,565 MSRP is 4.3x a 5090 for the same
bandwidth - you are paying for the memory, not speed. Street prices
above MSRP have been the norm since launch.
Run it locally: full CUDA stack - llama.cpp, Ollama, vLLM. The
modeldex results rank checkpoints by fit
and estimated tok/s.