TCPFN v2.1.1.50.3
MCP API REST API Performance Metrics

Performance

How long each tool takes, by the size of the panel it was given. Every number is a real, standalone call of the shipped tools on the hardware named — nothing reused from a previous call, nothing modelled or extrapolated.

measured on NVIDIA GeForce RTX 5080

At a glance

Per tool: how many runs were measured, the typical and slow-tail wall time, and the largest panel the tool has handled with the time that took.

627
runs measured
7
tools
1,915 × 254
largest panel, rows × columns
18.8 h
total wall time
ToolRunsMedian90th pct SlowestLargest panelTime on itHardware
discover1045.2 s16.5 s2.8 min1,915 × 2542.8 minNVIDIA GeForce RTX 5080
estimate_effect1042.8 s13.4 s2.3 min1,915 × 2542.3 minNVIDIA GeForce RTX 5080
is_identifiable1042.6 s13.0 s2.1 min1,915 × 2542.1 minNVIDIA GeForce RTX 5080
what_should_i_measure11.5 min1.5 min1.5 min200 × 201.5 minNVIDIA GeForce RTX 5080
rank_interventions1043.2 s13.7 s2.5 min1,915 × 2542.1 minNVIDIA GeForce RTX 5080
explain1043.2 s28.0 s31.1 min1,915 × 24231.1 minNVIDIA GeForce RTX 5080
causal_analysis10629.6 s3.5 min5.5 h1,915 × 2424.7 hNVIDIA GeForce RTX 5080

Size vs time

Each dot is one run; both axes are logarithmic. Time follows rows × columns: the widest panels are where the minutes go, and the same panel costs about the same on every call.

Every run

Filter by tool or by a minimum column count; click a column header to sort.

Tool Rows Columns Rows × columns Pairs Wall time Hardware

How to read the numbers

Rows and columns are what the tool actually analysed: the numeric columns left after preprocessing and the rows used, as reported in every response's diagnostics.data_report.
Wall time is the whole call as the response reports it in elapsed_seconds, including the necessary-condition checks and the stability re-check that run by default. For the tools that answer several pairs in one call the time is the sum over the pairs and the Pairs column says how many.
Hardware is the GPU the run was recorded on. A first call after start additionally pays the model load once.