community pool / model answer

Llama-3.3-70B-Instruct-NVFP4

chat · nvidia/Llama-3.3-70B-Instruct-NVFP4

23.7 tok/s · model median · 4 runs in the model record.

Answer

23.7tok/s

measuredCMP 170HX 64GB4-bit · 4 runs · decodebest cell in the pool

  1. CMP 170HX 64GB23.7 tok/smeasured4 runs · peak 57.4 GB · fits 64 GB rig memory
Reference rigQuantizationMedian resultBasisTTFT (ms)Peak VRAMMax contextEngines
CMP 170HX 64GB4-bit23.7 tok/smeasured6,70557.4 GB32,768vllm

Context tested

The table above is the tested context for each exact rig and quantization cell. Empty fields remain empty rather than being filled with an estimate.