Speed Bench

Speed Bench (report v3.1.0)

Web: speed-bench.html · JSON: field-exec-full-bench.json

Grok16 distro 3.0.0 · speed_demo suite v1.1.0 · schema v4 · 11 runners · 3s window

Version stamps

Field	Value
Report	3.1.0
Distro	3.0.0 (v3.0.0)
g16 pkg	Grok16-16.2.0
Bench suite	speed_demo @ 1.1.0
Schema	grok16-field-exec-full-bench/v4

Doctrine: data/grok16-speed-bench-version.json · Plate meld: data/grok16-plate-meld-bench-doctrine.json

Methodology

1. field-plate-meld.py fuse — fast plate fusion (24 plates, ~87 ms)
2. g16-compiler-sense-plate.py cycle — sense profile ladder
3. Compile every runner (gcc, g16 belt_1_0/2_0, sense expert, CMake)
4. Execute speed_demo @ 3s — axis field_execution_ops_per_sec
5. Post-meld re-exec same ELF — verify hot path
6. Cross-ref bench-all from data/bench/latest.json

Kernel

256×16 die · 240 frames/epoch · 512 prog_ops/frame · φ=0.618 · FieldX86 + entropy + wave + NEXUS

All executions (reference host @ 2026-06-27)

Runner	Profile	Compile	ops/s
CMake host g++ -O2	—	2,814 ms	85.8M
C++ host g++ -O2	—	1,710 ms	83.9M
C++ g16 sense expert	expert	1,608 ms	82.1M
C++ g16 belt_1_0	belt_1_0	1,888 ms	78.4M
C g16 belt_1_0	belt_1_0	395 ms	77.0M
C g16 belt_2_0	belt_2_0	296 ms	75.0M
C++ g16 belt_2_0	belt_2_0	2,021 ms	74.8M
C host gcc -O2	—	459 ms	73.7M
Python gpy-16	—	—	777K
Python CPython 3	—	—	748K

Plate meld verdict

Metric	Value
Meld generation	2
Plates fused	24
Sense profile	expert (eye_ear_green, score 0.75)
Sense vs belt_2_0 exec	+9.8% (82.1M vs 74.8M)
Sense vs belt_2_0 compile	−413 ms
Meld helps profile ladder	yes
Meld slows hot path	no

Professional conclusion: Plate meld helps situations where compiler-sense can pick a better profile than static belt_2_0. It does not replace compile-ahead for C/C++; it optimizes which profile wave-converts.

Reproduce

cd Grok16
export G16_PREFIX="$(pwd)"
export GROK16_SG_ROOT=/path/to/SG
export NEXUS_STATE_DIR=$GROK16_SG_ROOT/NewLatest/state
G16_PLATE_MELD_CMD=fuse SPEED_DEMO_TARGET_SEC=3 ./scripts/grok16-toolchain.sh exec-comprehensive-bench

Related

• Uncompiled Execution
• CMake and Linking
• Performance — belt triad + bench-all