Hardware Corner — GPU LLM Benchmarks per-GPU pages (llama.cpp, context sweep 4k-256k, March 2026)

Per-GPU pages under /gpu-llm-benchmarks/<gpu>/ with Prompt Processing and Token Generation tables by model and context tier (4k/16k/32k/64k/128k/256k). llama.cpp, GGUF Q4_K or MXFP4, single GPU, batch-1. Pages say Updated Benchmarks: March 2026. Some rows annotate a llama.cpp build (8189 on the 5090 Qwen3 8B and 30B-A3B rows, 8233 on the 5080 gpt-oss 20B row) and the 3090 Qwen3 rows annotate CUDA with -fa 1. Third-party in-lab measurement; raw llama-bench logs are not published, so confidence stays at 0.6.