Benchmark

Benchmark is built into Nexis. Its labelled title-bar launcher opens or focuses one dedicated companion window with a responsive setup board, run plan, live result matrix, and comparison charts. It is part of the ML Lab pack and the AI / ML and Everything presets.

Benchmark: pick models, engines and a workload, then run the comparison.

The historical standalone repository is archived; its Git history now lives in the main Nexis repository.

Backends

Backend	What is measured
ONNX Runtime	Real CPU inference through the linked ort runtime.
llama.cpp	Real GGUF inference through a llama-bench binary you locate.
nexis-ml-rs	Real training throughput on a standardized workload.
Simulated	Synthetic metrics for unsupported combinations and UI testing.

Every result carries a real or sim badge plus a note explaining what was actually measured. CPU-only ONNX is intentional; the current build does not advertise a GPU execution provider.

Workflow

1. Add .onnx or .gguf model files.
2. Select compatible backends and configure warm-up/measured runs.
3. Run the model × backend matrix.
4. Compare throughput, first-token latency, mean/p50/p95 latency, peak memory, and available accuracy metrics as cells stream in.

A run persists in the backend if the panel closes or reloads, and reopening the window reconnects to the active job rather than starting a second one. Completed results can be exported to CSV.

Shortcuts

Shortcuts are scoped to the Benchmark surface so they do not steal keystrokes from a terminal:

Key	Action
Ctrl+Enter (Command+Enter on macOS)	Run
Esc	Stop the active run

Theme switching belongs to Nexis; the former standalone t shortcut no longer exists.