Skip to content
Benchmarking & regression detection

Know a real speedup
from a lucky one.

A single benchmark number is a coin flip you didn't know you were taking. Kernalin measures the full picture and gives you a straight verdict — genuinely faster, no different, or slower — and blocks any slowdown before it reaches production. The wins you count are wins you can bank.

Request early access Explore the platform
kernalin · rope verdict 0 resamples
Headline reading
−32.7%
Honest verdict
Credibly faster
0.5× 0.75× 1.0× baseline 1.3×

99.4% of the posterior mass sits below the ROPE — convincingly beyond what this GPU's jitter could fake. This one earns the number.

Preview
Distributions, not points

A verdict, not a lucky run.

"12% faster" from one timing is a coin flip you didn't know you were taking. Kernalin builds a bootstrap posterior over the median-latency ratio and evaluates it with a Bayesian ROPE test: is the change credibly faster, statistically equivalent, or credibly regressed? The answer accounts for the noise instead of pretending it isn't there.
Where verdicts come from

Credibly faster

The posterior for the median-latency ratio sits convincingly beyond the noise floor.

Equivalent

The difference falls inside the region where the hardware genuinely can't tell them apart.

Regressed

Convincingly slower than baseline — flagged before it can ship.

Calibrated to your hardware

Every GPU gets its own noise floor.

A verdict is only as honest as its baseline. Kernalin calibrates each GPU to its own run-to-run jitter and stores baselines keyed to the environment that produced them — so comparisons are always apples-to-apples.

Distributions, not points

Every measurement is a full latency distribution. A single number throws away exactly the variance you need to make a decision.

Bayesian ROPE verdicts

A bootstrap posterior over the median-latency ratio, judged against a Region Of Practical Equivalence — not a fragile percentage threshold.

Per-GPU noise floor

Each GPU is calibrated for its own run-to-run jitter, so "faster" always means faster than this hardware's ability to fool you.

Environment-keyed baselines

Baselines are stored keyed to GPU, driver, and workload — you never compare across environments that aren't comparable.
Available
Regression gate

Four tiers between a change and main.

The same statistics that judge a speedup guard against a slowdown. The regression gate maps every metric diff to one of four outcomes and enforces it in CI — a credible regression on a gated metric simply cannot merge.
The verification gauntlet
BLOCK

A credible regression on a gated metric. The change cannot merge.

WARN

A suspected regression under the credibility threshold. Surfaced for a human.

INFO

A movement worth noting — within the noise band, no action required.

PASS

Equivalent or credibly faster. The change is clear to proceed.

The bottom line

Your benchmark shouldn't be able to lie to you.

Between distributional measurement, a per-GPU noise floor, environment-keyed baselines, and a Bayesian verdict, there is no gap left for a number to sneak through unearned.

It can't get lucky

A distribution over many trials leaves nowhere for a single fast run to hide.

It can't drift

Noise-floor calibration means "faster" is measured against this GPU's real variance.

It can't compare apples to oranges

Environment-keyed baselines refuse mismatched GPU, driver, or workload comparisons.

It can't sneak past CI

The tiered gate turns a credible regression into a hard BLOCK, not a footnote.

It can't hand-wave the math

A ROPE verdict is explicit about credibility — no silent 1% thresholds.

It can't be unauditable

Every verdict is backed by a stored, reproducible metric-diff.

Early access is open

Benchmarks you can gate on.

Bring distributional, regression-aware measurement to your GPU workloads. Join the early-access program.

Request early access Explore the platform