Processing time comparison between
onnxruntime and
onnx‑light‑cpu
(onnx‑light's reference evaluator with the SIMD-accelerated CPU kernels
registered) on the benchmark
examples
shipped by onnx‑light‑cpu, plus the benchmark-sized
big models exposed by onnx‑light. Each gallery
example exercises an operator over a range of inputs and shows
numpy as a baseline; each big model is a single larger graph
from onnx‑light's benchmark corpus. Backends run in separate global
phases with their default spin policies. Every measurement runs up to
CPU-scaled warm-up calls (not timed) and
up to CPU-scaled timed repetitions.
Each phase stops after 1 second of
cumulative execution.
speed-up (cpu) = onnxruntime /
onnx‑light‑cpu: values > 1 mean
onnx‑light‑cpu is faster than onnxruntime. Select an operator
summary row to expand its detailed measurements. The average speed-up is
sum(onnxruntime latency) / sum(onnx-light-cpu latency).
A measurement a backend could not produce shows failed;
hover it to read the recorded error.