RTX 3090 Ti llama.cpp: gpt-oss 20B (MXFP4) decode at 128k (Hardware Corner per-card page)

Hardware Corner per-card LLM benchmark page (March 2026 sweep). Same lab method as their RTX 5090 llama-bench series already ingested: llama.cpp llama-bench, Ubuntu 24.04, CUDA 12.8, Q4_K GGUF (gpt-oss 20B labeled MXFP4 on the page; filed under gguf-mxfp4-moe, the nearest schema format). Context depth 128k. 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; the sibling 5090 series used build 466c1911 with -fa 1 -ngl 99. That protocol is the likely one and is NOT asserted as measured on this card. Token generation 61.5 tok/s.

Transcribed from the Hardware Corner RTX 3090 Ti per-card page, not from the Dec 2025 ranking table (hardware-corner-gpu-ranking-local-llm). Where the two pages print the same cell they disagree — e.g. Qwen3 8B decode at 16k is 93.6 here and 126.94 on the ranking — so this record is status: contested rather than a replacement. 24 GB GDDR6X, 1,008 GB/s, 450 W. Page states Qwen3 32B Q4_K fits fully in VRAM at 16k context.