Burn 0.20 + CubeCL 0.9 on Intel Comet Lake GT2 (i5-10210U, UHD 620). Benchmark probe and findings.
- f16 GEMV finding and no-go decision — medium GEMV 5.72 → 2.80 ms, LM head 124.1 → 47.3 ms; both beat AVX2.
- vulkex bridge — the mistral.rs sidecar that measured the same hardware from the hand-written-executor side.
Tag: v0.1-f16-gemv
Status: educational; llama.cpp remains the production backend on this hardware. See the f16 finding for the reasoning.