Performance

Performance

Web: performance.html · speed-bench.html

Speed bench (5.1 — report v5.0.0)

Suite: speed_demo @ 1.1.0 · 3s execution window · Schema: grok16-field-exec-full-bench/v5

./scripts/grok16-toolchain.sh bench-refresh
# or stepwise:
SPEED_DEMO_TARGET_SEC=3 ./scripts/grok16-toolchain.sh exec-full-bench
./scripts/grok16-toolchain.sh exec-compare
./scripts/grok16-toolchain.sh bench-charts

Runner	Compile (ms)	ops/s	vs host
C++ g16 belt_2_0	BSP cache	102.8M	+1.7% vs g++
C++ host g++ -O2	BSP cache	101.1M	baseline
CMake host g++ -O2	BSP cache	98.9M	−2.2%
C g16 belt_1_0	BSP cache	99.5M	+2.1% vs gcc
C host gcc -O2	BSP cache	97.4M	baseline
C g16 belt_2_0	BSP cache	95.7M	−1.7%
Python host CPython 3	—	0.84M	baseline
Python gpy-16 GrokVM	—	0.76M	−10%

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

Artifacts: docs/SPEED-BENCH-REPORT.md · docs/field-exec-full-bench.json · Wiki: Speed-Bench

Overall (human read)

Lane	Verdict
C/C++ execution	Grok16 wins peak throughput on belt_2_0 (102.8M ops/s) and matches or beats host on C/C++ cold runs after BSP cache
Compile time	Host gcc/g++ still faster to compile; g16 pays compile cost for belt profiles
Triad runtime	belt_1_0 finishes the nexus bench in half the wall time of host g++ (2 ms vs 4 ms)
Python	Interpreter lane — host CPython still ahead on micro-bench; gpy-16 is uncompiled doctrine
Reliability	Self-monitor: 15/15 runs OK, zero drops, zero timeouts

Bottom line: Use Grok16 when you want sovereign compile + field belt execution at host parity or better on the FieldX86 kernel. Use field_physics for production FP truth; use belt_2_0 for peak ops/s.

Belt triad (2.0)

./scripts/grok16-toolchain.sh bench-triad
./scripts/grok16-toolchain.sh bench-charts

Workload: field-nexus-bench (240 frames, FieldX86 + entropy + wave + NEXUS). Host witness: gcc-14.

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

Artifact: data/bench/triad-latest.json · Chart: docs/assets/triad-chart.svg

bench-all profiles

./scripts/grok16-toolchain.sh bench-all

Profile	compile_ms	kernel wall_ms
belt_1_0	6305	2.89
belt_2_0	9962	3.66
field_opt	6337	3.08
field_physics	7636	4.06
vulkan_rtx	876	2.14

Artifact: data/bench/latest_field_opt.json · Chart: docs/assets/bench-all-chart.svg

Field vs host compare

./scripts/grok16-toolchain.sh bench-compare

Artifact: data/bench/compare-latest.json · Chart: docs/assets/compare-chart.svg

Related

• Speed Bench — full runner table + plate meld
• MCP — grok16_speed_bench tool for agents
• Profiles — belt_1_0, belt_2_0, field_opt, expert
• Release 5.1 — stack fabric + bench gates