Embedl-optimized
- timm/convnext_xxlarge.clip_laion2b_soup_ft_in1k
- INT8
- AGX Orin
- Labels
- Image
- Classification
- Build
- Embedl-optimized
- Precision
- INT8
- Runtime
- TensorRT 10.16.2.10 with Embedl optimizations
- Input size
- 256×256
convnext_xxlarge.clip_laion2b_soup_ft_in1k is an image classifier from timm. It labels an image with one of the 1,000 ImageNet classes.
Methodology
- Latency: trtexec on the NVIDIA Jetson AGX Orin: batch 1, 500 warm-up runs, then 200 timed; GPU compute is the engine's time on the GPU, without copies to and from it.
- Accuracy: the ImageNet validation set, 50,000 images, on the L4 build's INT8 engine. This page's engine is compiled from the same quantized ONNX, so its accuracy is the same up to the device's kernels, which the self-test checks.
- Self-test: during the build, on the AGX Orin: the runtime's own
embedl info,benchmark,inferand its KServe server, and on the L4 build's example images each engine's top class must be in the other's top 3 (they matched exactly on 3 of 3). - Build: the L4 build's quantized ONNX (Embedl Deploy, INT8 Q/DQ), compiled with TensorRT 10.16.2.10 on the device itself.
- Baseline TensorRT: the original model in FP16 on upstream TensorRT, same settings, same device.