RTX PRO 5000 Blackwell llama.cpp: Qwen3 30B-A3B (MoE) tg at 32k (Hardware Corner)

Hardware Corner llama.cpp llama-bench table, transcribed 2026-10-16 from the card page. Same lab method as the already-ingested RTX 5090 and RTX 6000 Ada series: llama.cpp, Q4_K or MXFP4 GGUF, per-request batch size 1. This is not an aggregate across concurrent requests. The page does not print the exact prompt and generation token counts for this cell; sibling 5090 rows in this repo were llama-bench runs with flash attention and full GPU offload, and that protocol is the likely one, not asserted as measured here. Context depth is the published column header. Context column: 32k.

Transcribed 2026-10-16 from the Hardware Corner context-scaling table for this card. Context depth 32k is the published column, not a separately logged prompt length. Per-request, batch 1, not aggregate.