Evidence, including the bad news

NAUX Benchmarks

This page publishes the latest local observation, its exact environment, and every reason it is not yet a competitive claim. Lower time is better.

4workloads

5measured implementations

30samples per row

3/20rows below 5% CV

Cross-language runtime snapshot

Median, tail latency, and variance

Raw JSON

Workload	Implementation	Median	p95	CV	Baseline / NAUX	Checksum
`sum_dense`Allocate, initialize, and repeatedly sum a dense integer list.	NAUX JIT	188.7 µs	254.1 µs	11.59%	1.000×	499950000
	C	196.9 µs	237.8 µs	7.35%	1.043×	499950000
	C++	79.9 µsfastest	90.5 µs	12.66%	0.423×	499950000
	Go	301.5 µs	373.3 µs	7.20%	1.598×	499950000
	Rust	99.0 µs	105.4 µs	2.92%	0.525×	499950000
`list_update`Sum and increment every element across repeated list passes.	NAUX JIT	210.5 µs	288.2 µs	12.27%	1.000×	500400000
	C	99.3 µsfastest	122.0 µs	10.91%	0.472×	500400000
	C++	100.8 µs	136.9 µs	11.25%	0.479×	500400000
	Go	108.7 µs	156.9 µs	17.94%	0.516×	500400000
	Rust	99.4 µs	131.8 µs	8.82%	0.472×	500400000
`dot_product`Accumulate value × value across a dense numeric list.	NAUX JIT	33.8 µsfastest	39.1 µs	5.68%	1.000×	3332833350000
	C	94.4 µs	99.5 µs	2.22%	2.789×	3332833350000
	C++	41.0 µs	48.8 µs	4.87%	1.212×	3332833350000
	Go	114.2 µs	157.4 µs	14.26%	3.375×	3332833350000
	Rust	96.9 µs	120.3 µs	7.04%	2.865×	3332833350000
`branch_mix`Run deterministic add/subtract recurrence through native branches.	NAUX JIT	1.394 ms	1.586 ms	5.93%	1.000×	-5151654
	C	99.5 µsfastest	120.5 µs	6.42%	0.071×	-5151654
	C++	99.7 µs	136.3 µs	16.92%	0.071×	-5151654
	Go	105.6 µs	172.5 µs	21.02%	0.076×	-5151654
	Rust	108.5 µs	150.3 µs	11.85%	0.078×	-5151654

Baseline/NAUX above 1.0 means NAUX was faster in this run; below 1.0 means the baseline was faster. All checksums matched. Ratios are observations, not admitted claims.

Measurement envelope

Environment

CPU: 11th Gen Intel Core i5-11400H @ 2.70 GHz
Host: Arch Linux 7.1.5 x86-64, glibc 2.44
Topology: 6 physical / 12 logical · 15.39 GiB
Control: core 0 pinned · powersave governor · turbo enabled
Workload: n=10,000 · reps=10 · 100 ms warmup
Repository: d31ee94a1771 · clean

No hidden VM fallback

NAUX execution certificate

Workload	Traces	Branches
sum_dense	1	2
list_update	1	2
dot_product	1	2
branch_mix	1	9

Every workload requested JIT, recorded one trace, and reported no fallback or deoptimization.

Repository performance gates

Regression evidence is mixed

Fixed-cost and cold-start gate

dot_runtime_n51214.5 µsPASS

dot_runtime_n102416.9 µsPASS

dot_runtime_n204819.1 µsPASS

dot_runtime_cold_n655361.490 msPASS

Scaling-slope gate

dot_runtime_only0.0590 ns/element · R² 0.9779FAIL

dot_trace_only0.0625 ns/element · R² 0.9997PASS

The runtime slope gate currently fails because its fit fell below the required R². The website does not collapse that failure into a green summary.

Reproduction

Run the same exploratory snapshot

The default large validation vector currently exceeds the learner work envelope, so this measured snapshot deliberately uses n=10000 and reps=10.

OUT_DIR=target/perf/cross_language \
CPU_CORE=0 N=10000 REPS=10 ITERS=30 WARMUP_MS=100 \
ENGINE=jit ENFORCE_CLAIM_ENV=0 ./scripts/bench_cross_language.sh

Performance contract Harness reference Benchmark sources Harness source