Add RKNN per-matmul apples-to-apples + mandate closed-runtime comparison - New page: per-matmul ork-driver vs closed RKNN matmul API (librknnrt). Same int8 (M,K,N) on the same NPU: decode ~1.24x (kernel efficiency), prefill ~3.2x (ork large-M tiling). Blog entry added (top). - Benchmark Standards: comparison against the closed runtime is now MANDATORY — rkllm (full-model tok/s) or rknn matmul API (per-matmul), whichever is applicable; match core count; added the rknn_vs_ork run command + refreshed baselines (183.9/11.71). - Submit-floor + run_multicore + open-vs-closed: added the measured RKNN per-matmul baseline (replaces the estimated rkllm-effective number).
Add Experiment Log blog + benchmark standards + 12 experiment writeups A running blog of NPU experiments across the open stack (ork-driver -> ggml-ork -> llama.cpp), most recent on top, linked from Home. Each full writeup carries the implementing code (commit + symbol + env-knob), the standardized benchmark (Qwen3-1.7B-w8a8, llama-bench warm, governors verified, vs librkllmrt on the same .rkllm), and a kept/reverted/ruled-out verdict. Plus the Benchmark Standards & Methodology page codifying the procedure and env knobs.