The table below reuses legacy numbers as synthetic inputs, not silicon data. A 12-way pilot lifts the L2 hit rate but slows every L2 hit and raises energy per instruction.
Net IPC falls on the measured mix. The pilot survives only where latency tolerance beats hit rate, segmented by SKU.
Metric | Baseline | 12-way pilot | Reading |
|---|---|---|---|
L2 hit rate | 87.1% | 90.4% | Plus 3.3 points of hits |
L2 hit latency | 13 cycles | 16 cycles | Plus 3 cycles on every L2 hit |
LLC MPKI | 11.9 | 10.7 | Minus 1.2 misses per kilo-instruction |
Energy per instruction | 1.82 nJ | 1.98 nJ | Plus 0.16 nJ per instruction |
IPC | 1.46 | 1.44 | Minus 0.02, so the pilot loses here |
One method handles every miss regression. Split the MPKI delta into demand, prefetch-evicted and coherence shares before touching any structure.
Freeze counters and windows: confirm the same events and sampling on both sides.
Split the delta: demand misses, prefetch-evicted lines and coherence invalidations.
Normalize and localize: convert with accesses per kilo-instruction, then map hot sets with a pressure heatmap.
Classify: dead-on-fill means pollution or insertion trouble, hot sets mean conflict or policy, flat spread means capacity or workload shift.
Run one bounded experiment and sign off with latency, bandwidth and fairness guards.
synthetic regression: LLC MPKI 18.4 -> 24.1, delta +5.7
demand +3.1 (54.4% of the delta)
prefetch +2.2 (38.6% of the delta)
coherence +0.4 (7.0% of the delta)
shares sum to 100.0% up to rounding
top-42 hot sets hold about 60% of misses with high dead-on-fill share
verdict: insertion and prefetch admission suspect, not capacityCapacity pressure or a workload footprint shift. Neither conflict fixes nor insertion tuning match a flat spread.
Tighten prefetch admission or depth on a shadow run, and watch demand hits plus p95 latency before any rollout.