Hardware Corner RTX PRO 5000 Blackwell local LLM benchmark page

Hardware Corner GPU LLM benchmark page, page header says Updated Benchmarks: March 2026. Same lab methodology as the RTX 5090 page already ingested (Allan Witt, llama-bench, Ubuntu 24.04, CUDA 12.8, Q4_K / Q4_K_XL GGUF, per-request batch-1, context-depth sweeps). Headline on the page: 84.3 t/s generation and 3,064 t/s prompt processing on 14B 4-bit at 16k context. Full tables transcribed for Qwen3 8B, Qwen3 14B, gpt-oss 20B MXFP4, Qwen3 30B A3B, Qwen3 32B, Llama 3.3 70B at 4k through 256k where the page prints a number. Page also states the card runs Llama 3.3 70B Q4_K fully in VRAM (48 GB). Not a vendor measurement.