Quality + performance
A faster wrong answer still fails.
Every performance result keeps model quality, device context, and runtime version beside it.
Best validated P5028.4msONNX Runtime · FP32
Quality delta0.0ppAgainst PyTorch baseline
Peak memory saved174 MBValidated runtime
Device contextCPUApple silicon development host
Benchmark registry
Inference runs
| Model | Runtime | Precision | Quality | P50 | P95 | Peak memory | Status |
|---|---|---|---|---|---|---|---|
| ResNet-18 | PyTorch | FP32 | 89.8% | 42.6ms | 51.2ms | 612 MB | Baseline |
| ResNet-18 | ONNX Runtime | FP32 | 89.8% | 28.4ms | 34.7ms | 438 MB | Validated |
| ResNet-18 | TensorRT | FP16 | 89.6% | 8.9ms | 11.3ms | 326 MB | Phase 2 target |