HOSTKEY: RTX PRO 5000 72GB Blackwell, gpt-oss:20b, Ollama, avg decode

Per-request decode tok/s, one request at a time (the table reports a single average tokens/sec, not an aggregate across concurrent requests). HOSTKEY own-hardware test, article dated 2026-08-09, of one NVIDIA RTX PRO 5000 Blackwell 72 GB (PCI id 10de:2bb3, double-sided memory) in a Ryzen 9 7950X host, 128 GB DDR5, Ubuntu 24.04, Ollama. Model is the Ollama tag gpt-oss:20b. gpt-oss-20b pulled as the Ollama tag gpt-oss:20b; the page flags it as a mixture-of-experts model. Ollama version, the GGUF quantization tag, the prompt length and the number of generated tokens are NOT stated on the page, so format_id is null and no prompt or output length is asserted. The companion R9700 article (hostkey.com/blog/142) describes the same harness: a custom Ollama test reporting average tokens/sec, max context, load time and generation time.

First published measurement found for nvidia-rtx-pro-5000-72gb-blackwell, which previously had zero benchmark records. The page also prints a max-context column and a load-time column; those are not filed because neither is a metric this record type measures. Do not read this number as comparable to the Hardware Corner llama-bench tables for the 48 GB SKU: different engine, unstated quant, unstated context depth.