RTX PRO 5000 Blackwell llama.cpp: Qwen3 30B-A3B (MoE) decode at 16k (Hardware Corner)

Hardware Corner RTX PRO 5000 Blackwell LLM benchmark page (updated March 2026). Same lab method as their RTX 5090 llama-bench series already ingested: llama.cpp llama-bench, Ubuntu 24.04, CUDA 12.8, Q4_K / Q4_K_XL GGUF (gpt-oss 20B labeled MXFP4 on the page). Context depth 16k. Per-request, batch size 1, NOT aggregate across concurrent requests. The page does not print the llama.cpp commit or the exact prompt/output lengths for this card; the sibling 5090 series in this repo used build 466c1911 with -fa 1 -ngl 99, prompt 1024/2048, output 128. That protocol is the likely one and is NOT asserted as measured on this card.

Transcribed from the March 2026 Hardware Corner table for the 48 GB RTX PRO 5000 Blackwell. The 72 GB SKU has no public row on this page and is a separate accelerator id. Page headline 84.3 t/s generation and 3,064 t/s prompt processing for 14B 4-bit at 16k matches the 14B 16k cells. 48 GB GDDR7; Llama 3.3 70B Q4_K reported to fit fully in VRAM.