AMD Radeon Pro W7900 48GB — Benchmarks & Specs

Bottom line: how fast is the AMD Radeon Pro W7900 48GB?

SpecPicks median: 108 tokens/sec across 4 cited single-user runs of 7-9B dense models at 4-bit (Q4) under llama.cpp (range 68.5–127 tok/s), from llama.cpp GitHub (2 runs), llm-tracker.info (2 runs). For larger models, the SpecPicks median is 11.3 tokens/sec across 7 cited single-user runs of 65-73B dense models at 4-bit (Q4) under llama.cpp (range 10.4–12.7 tok/s), from Tom's Hardware (3 runs), Windows Forum, llm-tracker.info (3 runs). For local LLM inference it generates 11.3 tokens/sec running llama3.1:70b at q4_K_M under llama.cpp, per llm-tracker.info. In PassMark G3D Mark it scores 27,729 points, per PassMark. Its 48 GB of VRAM is the binding constraint for local inference: the llama3.1:70b run above (70.6B parameters) is the largest model on file at 4-bit quantization on this card. That run's source does not report a context length; the Will-It-Run matrix sizes each model at its listed quantization with an ~8K context, and by that measure DeepSeek-R1-Distill-Llama 70B at Q4_K_M is the largest that fits in 48 GB.

Every figure above is a row in the tables below, and each row links out to the review or public benchmark database the number was taken from; the SpecPicks median is computed here from the runs it names, each a row in the AI table. SpecPicks aggregates published measurements; it does not report first-party benchmark runs.