Speed bench — report v5.0.0

Distro 5.0.0 · suite speed_demo @ 1.1.0 · schema grok16-field-exec-full-bench/v5

Version stamps: Every run writes docs/field-exec-full-bench.json with distro, g16, suite, and report versions. Doctrine: data/grok16-speed-bench-version.json

Speed bench bar chart — Execution ops/s by runner — 2026-06-27T19:20:27Z on `default-X870-Pro-RS`

Winners

Fastest execution102.84MC++ — g16 belt_2_0

Fastest compile— ms—

Best Python843,829interpreter — no compile

Amortized first-run102.84MC++ — g16 belt_2_0

Full results (3s window)

Runner	Compile (ms)	Exec wall (ms)	ops/s	Binary bytes
C++ — g16 belt_2_0	—	3,071	102,843,184.74	22,912
C++ — host g++ -O2	—	3,083	101,093,393.65	21,640
C — g16 belt_1_0	—	3,006	99,454,820.63	16,192
CMake — host g++ -O2	—	3,014	98,892,364.74	21,640
C — host gcc -O2	—	3,006	97,416,875.40	24,568
C — g16 belt_2_0	—	3,006	95,729,447.71	16,248
C++ — g16 belt_1_0	—	3,040	95,052,367.49	22,784
C++ — g16 belt_2_0 (post-meld re-exec)	—	3,008	93,388,093.30	22,912
C++ — g16 sense belt_2_0	—	3,078	91,070,985.42	22,784
Python — host CPython 3	—	3,110	843,828.62	—
Python — gpy-16 GrokVM	—	3,210	755,552.29	—

Plate meld analysis

Cycle: field-plate-meld.py fuse + g16-compiler-sense-plate.py cycle before compiles. Doctrine: data/grok16-plate-meld-bench-doctrine.json

Metric	Value
Meld generation	0
Plates fused	0
Compiler sense profile	`belt_2_0` (unknown)
Sense vs belt_2_0 exec ratio	0.8855
Meld helps profile ladder	no

Professional verdict: Plate meld does not block the ELF hot path. On this host, compiler-sense compile was 0 ms vs static belt_2_0; sense execution ratio 0.8855. Profile ladder unlocks when meld generation ≥ 0.

bench-all cross-reference

10 profile runs from data/bench/latest.json (field-nexus-bench). Run: ./scripts/grok16-toolchain.sh bench-all

Doctrine

Dev / uncompiled: Python at interpreter speed (~0.75–0.78M ops/s). C/C++ rely on chamber compile ahead.
Plate meld: Fuses sense plates → compiler-sense profile ladder before wave-convert.
Compare axis: field_execution_ops_per_sec on identical speed_demo kernel.

See Uncompiled execution · CMake & linking

Reproduce (comprehensive)

cd Grok16
git checkout v5.2.0
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
./scripts/grok16-toolchain.sh exec-compare

JSON: field-exec-full-bench.json · MD: SPEED-BENCH-REPORT.md