RTX PRO 6000 Blackwell is the desktop version of the B200's brain. Same Blackwell architecture, same FP4 support, but built for a workstation: 96GB of GDDR7 ECC, a single-slot-ish card you can put in a tower, and no NVLink between cards. It's the answer to "what do I run when I don't have a rack?"
The numbers
- Memory: 96GB GDDR7 ECC at ~1.8TB/s (workstation edition). The 96GB is the headline: it fits a 70B at 4-bit, or a 405B at 4-bit across two cards.
- FP4: ~4000 TFLOPS dense (Blackwell's 5th-gen tensor cores with micro-tensor scaling).
- FP8: ~2000 TFLOPS dense.
- FP16: ~1000 TFLOPS dense.
- Power: 600W TDP. Needs a beefy PSU, but no liquid cooling.
- Price: $15,499 MSRP (Workstation Edition).
Measured, not marketing
From local.ai's independent benchmarks (llama.cpp and vLLM, real task runs):
- LFM2.5 8B A1B (Q8_0): 460 tok/s, τ² 94.8%
- LFM2.5 8B A1B (default, vLLM): 389 tok/s
- Gemma 4 E2B It (UD-Q4_K_XL): 300 tok/s
- Nemotron 3 Nano Omni 30B A3B Reasoning (nvfp4): 298 tok/s
- Holo 3.1 35B A3B (nvfp4): 279 tok/s
- Gpt Oss 20B (mxfp4): 271 tok/s
- Holo 3.1 35B A3B (q4_k_m): 271 tok/s
- Ornith 1.0 35B (Q4_K_M): 270 tok/s
- Laguna XS 2.1 (Q4_K_M): 262 tok/s
- Qwen3.6 35B A3B (UD-Q2_K_XL): 255 tok/s
That's 54 measured benchmarks on this card. The pattern: small MoE models fly (460 tok/s on an 8B), and a 35B A3B MoE stays interactive (255-279 tok/s) even in 4-bit. For a personal playground, this is the top of the consumer stack.
RTX 6000 Ada: the previous generation
If you're buying used or want ECC at half the price, the RTX 6000 Ada (48GB GDDR6 ECC, ~$8,219) is still a beast. It's Ada architecture (no FP4, but FP8 via 4th-gen tensor cores):
- LFM2.5 8B A1B (Q8_0): 311 tok/s
- Gemma 4 E2B It (UD-Q4_K_XL): 239 tok/s
- Nemotron 3 Nano 30B A3B (nvfp4): 218 tok/s
- Qwen3.6 35B A3B (nvfp4): 217 tok/s
- Holo 3.1 35B A3B (q4_k_m): 188 tok/s
- Ornith 1.0 35B (Q4_K_M): 188 tok/s
The 48GB is the real constraint: a 70B at 4-bit needs ~40GB, so it fits, but a 405B needs 8 of them. The 96GB of the PRO 6000 is the unlock.
Why it matters for inference
These cards are the "personal playground" tier. They let you test quantization schemes, run a small vLLM instance, and prototype kernels without touching a cluster. The FP4 support means the weights you serve in production (NVFP4) are the same format you can test locally. That's the real value: format parity between your desk and your data center.
Sources
- local.ai hardware benchmarks: independent measured tok/s for these cards.
- NVIDIA RTX PRO 6000 product page: official specs.
- NVIDIA RTX 6000 Ada product page: official specs.