cache-roof — kernel grading lab

Machine peaks (editable)

Peak compute (GFLOP/s) DRAM bandwidth (GB/s)

L2 bandwidth (GB/s, ≈4× DRAM) CI half-width (± %, synthetic demo band; real harness measures it)

L2-hit rate 65% (0–95%)

Dequant-overlap q 35% (0–90%)

Hit splits each kernel's bytes: B_dram=B·(1−h), B_L2=B·h. Overlap hides fraction q of memory time (quantized kernels).

Kernels — edit FLOPs, bytes, achieved (GFLOP/s)

kernel	FLOP (GF)	bytes (GB)	achieved	roof	grade %	bound	verdict

Verdict thresholds mirror roofline.analyze: ≥80% at-roof · ≥50% moderate · ≥20% significant · else overhead-bound.

Ladder — % of cache-adjusted peak (CI whiskers)

Formulas ported from cache_roof/roofline.py: cache_adjusted_roof + analyze. Bound = compute if t_comp ≥ t_eff, else cache if B_L2>B_dram, else memory. Classic DRAM-only roof shown in tooltip column via title attribute.