Hardware Corner Mar-2026: nvidia-rtx-3090 llama.cpp gpt-oss-20B MXFP4 GGUF token generation 147.5 tok/s at 4k context

THIRD-PARTY MEASURED (hardware-corner.net /gpu-llm-benchmarks/rtx-3090/, page updated March 2026, accessed 2026-10-16). llama.cpp llama-bench, GGUF, single GPU, full offload. The page publishes prompt-processing tok/s and token-generation tok/s by context tier; it does not state the llama.cpp build, the prompt length, or the generation length. The page annotates this model row CUDA, -fa 1 (flash attention on). Value is the 4k context column. Per-request batch-1, NOT aggregate across concurrent requests.

Context tier is the page bucket, not an exact measured token count. Page quant Q4_K maps to gguf-q4-k-m; page MXFP4 maps to gguf-mxfp4-moe. Distinct from the older Q4_K_XL rows sourced from hardware-corner-gpu-ranking-local-llm.