JetPack 4.3 Developer Preview
JetPack 4.3 Developer Preview provides an early look at two JetPack 4.3 components: TensorRT 6.0.1 and cuDNN 7.6.3.
Only TensorRT, cuDNN, CUDA, and L4T are included in this release. Only Jetson AGX Xavier Developer Kit is supported.
JetPack Component |
JetPack 4.3 Developer Preview |
JetPack 4.2.2 |
---|---|---|
L4T | 32.2.2 | 32.2.1 |
MM API | not available | 32.2.1 |
Container Runtime | not available | 0.9.0 beta 1 |
CUDA | 10.0.326 | 10.0.326 |
cuDNN |
7.6.3 |
7.5.0.56 |
TensorRT |
6.0.1 |
5.1.6.1 |
VisionWorks | not available | 1.6 |
OpenCV | not available | 3.3.1 |
Host computer components | not available |
CUDA, Computer Vision, and Developer Tools |
The production release of JetPack 4.3 will support all Jetson products and include all JetPack components.
JetPack 4.2.2 remains the latest production release, supporting all Jetson products
Installing JetPack 4.3 DP:
Download JetPack 4.3 Developer Preview and follow the command line instructions below:
NVIDIA SDK Manager does not support installation of JetPack 4.3 Developer Preview. Follow these command line instructions to flash your Jetson AGX Xavier Developer Kit and install CUDA, cuDNN, and TensorRT.
Key Features in JetPack 4.3 Developer Preview
JetPack 4.3 DP introduces TensorRT 6.0.1 and cuDNN 7.6.3.OS
NVIDIA L4T provides the bootloader, Linux kernel, necessary firmwares, NVIDIA drivers, sample filesystem, and more.
JetPack 4.3 DP Highlights:
-
L4T 32.2.2 features are unchanged from L4T 32.2.1 (included in JetPack 4.2.2) except:
- support is added for TensorRT 6.0.1 and cuDNN 7.6.3
- only Jetson AGX Xavier Developer Kit is supported
For details, see L4T 32.2.2 release notes.
TensorRT
TensorRT is a high performance deep learning inference runtime for image classification, segmentation, and object detection neural networks. It speeds up deep learning inference as well as reducing the runtime memory footprint for convolutional and deconv neural networks.
JetPack 4.3 DP Highlights:
- New layers, operators, and resize operations
- New samples
- New optimizations
For details, see TensorRT 6.0.1 for Jetson platform release notes.
cuDNN
CUDA Deep Neural Network library provides high-performance primitives for deep learning frameworks. It includes support for convolutions, activation functions and tensor transformations.
JetPack 4.3 DP Highlights:
- Tensor Core Support and performance enhancements for 3D convolutions, grouped convolutions, and more
- Performance enhancements for 3D deconvolution
- Enhanced functionality for fused operations
For details, see cuDNN 7.6.3 release notes.