torchvision/alexnet

Methodology

• Latency: on the NVIDIA L4, batch 1, median GPU compute time (no copies to and from the GPU), from the build's own benchmark.
• Accuracy: the ImageNet-1k validation split, 50,000 images, from the build's own validation on the NVIDIA L4.
• Baseline TensorRT: the original model in FP16 on upstream TensorRT, same settings, same device. Not measured yet.
• Build: Embedl Deploy quantization (INT8 weights per channel and activations), calibrated with ImageNetCalibDataset, 5000 batches of 1, exported to ONNX and compiled with Embedl's TensorRT transforms.