Profiling results

Call result.to_dataframe() on the ProfileResults returned by nsight.analyze.kernel() to inspect the processed results. Each row summarizes one annotation, metric, and configuration across repeated runs. Function parameters appear as configuration columns, such as n in the Quickstart.

Measurement and statistics columns

Column	Meaning
Annotation, Metric	Annotated region and collected or derived metric.
AvgValue	Arithmetic mean of the valid metric samples, or the ratio to a baseline when normalization is requested (see below).
StdDev	Sample standard deviation, using a denominator of NumRuns - 1.
MinValue, MaxValue	Minimum and maximum valid sample values.
NumRuns	Number of non-missing metric samples used in the aggregation. This can be smaller than the requested runs when samples are missing.
CI95_Lower, CI95_Upper	Approximate 95% confidence bounds, computed as AvgValue +/- 1.96 * StdDev / sqrt(NumRuns) before normalization. These use a normal approximation, including for small sample counts.
RelativeStdDevPct	100 * StdDev / abs(AvgValue), computed before normalization. It is NaN when the mean is zero or the standard deviation is undefined.
StableMeasurement	True exactly when RelativeStdDevPct < 2.0. This is a variability heuristic, not a guarantee of benchmark accuracy. Undefined relative variability yields pd.NA (displayed as <NA>), the missing value in a pandas nullable Boolean column.
Kernel, GPU, Host	Kernel name and execution environment.
ComputeClock, MemoryClock, Unit	Clock metadata and the original metric unit. Availability depends on the profiling backend.

One run and missing samples

The default is runs=1. One valid sample is enough to report a value, but not enough to estimate sample standard deviation. StdDev, RelativeStdDevPct, and both confidence bounds are therefore NaN. StableMeasurement is pd.NA in this case, distinguishing unknown stability from an observed relative standard deviation at or above 2%.

Use @nsight.analyze.kernel(runs=10) to collect repeated measurements. At least two valid samples are needed for sample standard deviation. Even with repeated runs, a zero mean leaves RelativeStdDevPct undefined. Missing samples are excluded from the count and standard aggregations; a group with no valid samples has NumRuns=0 and undefined statistics.

Normalization and geometric mean

Pass normalize_against="baseline" to nsight.analyze.kernel() to normalize against an annotation named baseline for the same metric and configuration.

Column	Meaning
Normalized	False when normalization is disabled; otherwise the name of the baseline annotation. It is not always a Boolean.
NormalizationValue	Baseline’s unnormalized mean. Present only when normalization is enabled.
Geomean	Geometric mean of AvgValue across configurations for the same annotation and metric, repeated on each row of that group. It is computed after normalization, if enabled. With one positive value, it equals that value. Missing values are excluded; negative values are outside the real-valued geometric mean’s domain.

For example, an unnormalized mean of 20 ns and a baseline mean of 10 ns produce AvgValue=2.0, NormalizationValue=10.0, and Normalized="baseline". The ratio is dimensionless. The current implementation normalizes only AvgValue: StdDev, MinValue, MaxValue, confidence bounds, and Unit retain their original scale or metadata. RelativeStdDevPct and StableMeasurement describe the unnormalized samples. Do not interpret the original confidence bounds as bounds on the normalized ratio.

Custom information collectors can add further columns. See nsight.analyze for the experimental collector API.