One synthetic score hides the mix. Three workload classes with stated weights replace it here.
Class | Speedup vs baseline | Weight |
|---|---|---|
Sustained compute | 1.30x | 0.5 |
Burst responsiveness | 1.10x | 0.3 |
Background tasks | 1.05x | 0.2 |
Ratios combine with a weighted geometric mean. Weights sum to 1.0, and all inputs are illustrative.
Weighted geomean (illustrative speedups and weights)
formula = exp(0.5 x ln(1.30) + 0.3 x ln(1.10) + 0.2 x ln(1.05))
ln(1.30) = 0.26237 -> x 0.5 = 0.13118
ln(1.10) = 0.09531 -> x 0.3 = 0.02859
ln(1.05) = 0.04879 -> x 0.2 = 0.00976
sum = 0.13118 + 0.02859 + 0.00976 = 0.16953
exp(0.16953) = 1.18x
plain mean = 0.5 x 1.30 + 0.3 x 1.10 + 0.2 x 1.05 = 1.19xThe plain mean reads 1.19x and the geomean reads 1.18x. Small gaps compound when targets chain across releases, so the ledger records the geomean.
Power without throughput is half a metric. Energy per operation divides the two under stated assumptions.
Energy per operation (illustrative)
assume P = 1.8 W sustained, throughput = 3.0 x 10^9 ops/s
E = P / throughput = 1.8 / 3.0e9 = 0.6e-9 J = 0.60 nJ/opA faster block that burns proportionally more power shows no energy gain. Report both numbers or the tradeoff stays invisible.
Guardbands add, they do not overlap. Each component below is an illustrative assumption in picoseconds.
Timing guardband stack (illustrative, 1 GHz clock = 1000 ps period)
variation + aging + modeling = 30 + 25 + 25 = 80 ps
share of period = 80 / 1000 = 8%The geomean is 1.18x and the plain mean is 1.19x, as worked above.
0.60 nJ/op, from 1.8 divided by 3.0e9.