Speed Bench

Speed Bench v5.1.0

Live JSON: docs/field-exec-full-bench.json · triad: data/bench/triad-latest.json

Host: Linux x86_64 · g16 @ 16.2.0 · suite speed_demo @ 1.1.0 · 3 s window

Winners (measured 2026-06-27)

Category	Result
Fastest execution	102.8M ops/s — C++ g16 belt_2_0
C++ host g++ -O2	101.1M ops/s
C belt_1_0	99.5M ops/s
C host gcc -O2	97.4M ops/s
Python host CPython	0.84M ops/s (interpreter — no compile)
Python gpy-16	0.76M ops/s
Self-monitor	15 runs, 0 dropped, 0 timeouts

Overall: Grok16 matches or beats host gcc/g++ on the FieldX86 kernel at steady state. Compile time is higher; execution throughput peaks on belt_2_0.

Belt triad (bench-triad)

Workload: field-nexus-bench — 240 frames, FieldX86 + entropy + wave + NEXUS.
Updated: 2026-06-29

Toolchain	Profile	compile_ms	run wall_ms	binary bytes
host g++	-O3 -march=native	2078	4	27264
g16	belt_1_0	4961	2	22696
g16	belt_2_0	5068	3	22824

belt_1_0 runs 2× faster wall time than host on this triad workload.

bench-all profiles (kernel wall)

Profile	compile_ms	run_ms	kernel
field_opt	6337	7	wall ~3.08 ms
belt_1_0	6305	6	wall ~2.89 ms
belt_2_0	9962	11	wall ~3.66 ms
field_physics	7636	13	wall ~4.06 ms · safe FP
vulkan_rtx	876	4	wall ~2.14 ms

nexus_checksum may be -nan under -ffast-math — use field_physics when checksum truth matters.

Reproduce

export G16_PREFIX="$(pwd)"
./scripts/grok16-toolchain.sh bench-triad
./scripts/grok16-toolchain.sh bench-refresh
./scripts/grok16-toolchain.sh exec-full-bench

Via MCP: grok16_speed_bench · grok16_toolchain exec-full-bench

Charts: docs/assets/speed-bench-chart.svg, triad-chart.svg, compare-chart.svg

Web: Speed Bench manual