Zinc Next: M4 benchmarks

Median of 11 runs (one warm-up), wall clock, ms. Machine: macOS-15.7.4-arm64-arm-64bit. Interpreter: zinc run of the compiled bytecode; AOT: zinc build; QuickJS: qjs on the stripped kernel; native: the old toolchain’s release build. tools/bench-m4 regenerates this file and bench/m4.json.

kernel	interp	AOT	QuickJS	native	AOT / native	QuickJS / interp
fib	45.3	21.0	247.2	14.1	1.50x	5.46x
nbody	485.5	111.9	2637.7	39.2	2.85x	5.43x
mandelbrot	103.5	31.3	700.9	25.7	1.22x	6.77x
spectralnorm	764.3	66.6	4110.3	57.0	1.17x	5.38x
fannkuchredux	1161.4	538.4	7480.6	367.2	1.47x	6.44x
binarytrees	33.8	19.4	109.9	17.4	1.12x	3.25x
sort	137.6	105.3	814.6	89.5	1.18x	5.92x
mapset	52.7	24.4	115.9	15.0	1.63x	2.20x
strings	44.3	36.8	59.5	31.1	1.18x	1.34x
jsonout	15.8	10.2	7.1	6.4	1.59x	0.45x
dynsum	39.2	29.4	30.2	n/a	n/a	0.77x

Thresholds

kernel	threshold	measured	result
fib	AOT within 3x of native	1.50x	pass
fib	interpreter at least 5x QuickJS	5.46x	pass
nbody	AOT within 3x of native	2.85x	pass
nbody	interpreter at least 5x QuickJS	5.43x	pass
mandelbrot	interpreter at least 5x QuickJS	6.77x	pass
spectralnorm	interpreter at least 5x QuickJS	5.38x	pass
fannkuchredux	interpreter at least 5x QuickJS	6.44x	pass
binarytrees	AOT within 3x of native	1.12x	pass
sort	AOT within 3x of native	1.18x	pass
strings	interpreter not slower than QuickJS	1.34x	pass
jsonout	interpreter not slower than QuickJS	0.45x	LOSS
dynsum	interpreter not slower than QuickJS	0.77x	LOSS

Resources (one extra run per cell)

kernel	interp RSS MB	AOT RSS MB	interp user s	AOT user s
fib	16.62	9.45	0.04	0.01
nbody	17.08	9.73	0.47	0.10
mandelbrot	16.56	9.44	0.09	0.02
spectralnorm	16.89	9.78	0.75	0.06
fannkuchredux	16.75	9.58	1.15	0.53
binarytrees	17.52	10.11	0.02	0.01
sort	40.25	33.14	0.12	0.09
mapset	22.00	14.75	0.05	0.02
strings	38.34	31.16	0.03	0.03
jsonout	17.20	10.12	0.01	0.00
dynsum	18.97	11.83	0.03	0.02

Edit this page on GitHub ↗