Embedl-optimized
- facebook/deit-base-patch16-224
- FP8
- AGX Thor
- Labels
- Image
- Classification
- Build
- Embedl-optimized
- Precision
- FP8
- Runtime
- TensorRT 10.13.3.9 with Embedl optimizations
- Input size
- 224×224
DeiT-Base is Facebook's data-efficient image transformer for image classification: it labels an image with one of the 1,000 ImageNet classes.
Methodology
- Latency: trtexec on the NVIDIA Jetson AGX Thor with TensorRT 10.13.3.9: batch 1, CUDA graphs, median GPU compute time (no copies to and from the GPU).
- Accuracy: the ImageNette validation split, 3,925 images with ImageNet-1k labels, on both engines.
- Memory: peak device memory while the engine runs.
- Size: the serialized TensorRT engine.
- Baseline TensorRT: the original model in FP16 on upstream TensorRT 10.13.3.9, same settings, same device.
- Build: Embedl Deploy fusion and quantization, calibrated on 128 ImageNette training images, exported to ONNX and compiled with trtexec.