NVIDIA Jetson Orin - JetPack Components
This section provides an overview of some of the main software components provided in Jetpack 5.0 for the Orin AGX.
Jetpack Components
Table 1 provides the versions of the key packages included in Jetpack 5.0, and the version previously available in Jetpack 4.6.1:
Package | Jetpack 5.0 Version | Jetpack 4.6.1 Version |
---|---|---|
Linux for Tegra | 34.1 | 32.7.1 |
CUDA | 11.4 | 10.2 |
cuDNN | 8.3.2 | 8.2.1 |
TensorRT | 8.4.0 Early Access | 8.2.1 |
OpenCV | 4.5.4 (Developer Preview) | 4.1.1 |
Vision Programing Interface (VPI) | 2.0 | 1.2 |
Linux for Tegra
Linux for Tegra (L4T) is the component that provides the OS components (kernel, file system, drivers, firmware) and the bootloader. Jetpack 5.0 comes with L4T version 34.1, featuring:
- Linux Kernel version 5.10: the previous Jetpack version (4.6) came with Linux Kernel version 4.9.
- Ubuntu 20.04-based root file system: the previous Jetpack version (4.6) came with Ubuntu 18.04-based root file system.
- UEFI-based bootloader: CBoot was used prior to L4T 34.1.
- Support for NVSCI: Provides utilities for streaming data packets between different applications and for inter-process communication (IPC).
- Toolchain: aarch64 gcc-9.3-glibc-2.31 for 64-bit Kernel and user space.
- Jetson Power GUI: a new tool for monitoring the power and thermal status of the Jetson platform.
Multimedia API
Included as part of L4T, the Multimedia API allows developers to develop camera-based multimedia applications. It has two main components:
- V4L2 Sensor driver API: the V4L2 API that enables the capture devices, codecs, and scaling functionality.
- Libargus: the proprietary NVIDIA library that allows using the Jetson Image Signal Processor (ISP).
JetPack 5.0 Multimedia API features new NVUtils Multimedia buffer management and buffer transformation API. Previous nvbuf_utils is deprecated in this release.
CUDA
Jetpack 5.0 includes CUDA 11.4 support. CUDA is a parallel computing API for NVIDIA GPUs. NVIDIA's CUDA Toolkit provides the development environment for creating GPU-accelerated applications in C/C++. Starting with CUDA 11, the various components in the toolkit are versioned independently. Table 2 provides the CUDA 11.4 components version information:
Component Name | Version Information | Supported Architectures |
---|---|---|
CUDA Runtime (cudart) | 11.4.148 | x86_64, POWER, Arm64 |
cuobjdump | 11.4.120 | x86_64, POWER, Arm64 |
CUPTI | 11.4.120 | x86_64, POWER, Arm64 |
CUDA cuxxfilt (demangler) | 11.4.120 | x86_64, POWER, Arm64 |
CUDA Demo Suite | 11.4.100 | x86_64 |
CUDA GDB | 11.4.120 | x86_64, POWER, Arm64 |
CUDA Memcheck | 11.4.120 | x86_64, POWER |
CUDA NVCC | 11.4.152 | x86_64, POWER, Arm64 |
CUDA nvdisasm | 11.4.152 | x86_64, POWER, Arm64 |
CUDA NVML Headers | 11.4.120 | x86_64, POWER, Arm64 |
CUDA nvprof | 11.4.120 | x86_64, POWER, Arm64 |
CUDA nvprune | 11.4.120 | x86_64, POWER, Arm64 |
CUDA NVRTC | 11.4.152 | x86_64, POWER, Arm64 |
CUDA NVTX | 11.4.120 | x86_64, POWER, Arm64 |
CUDA NVVP | 11.4.193 | x86_64, POWER |
CUDA Samples | 11.4.120 | x86_64, POWER, Arm64 |
CUDA Compute Sanitizer API | 11.4.120 | x86_64, POWER, Arm64 |
CUDA Thrust | 11.4.122 | x86_64, POWER, Arm64 |
CUDA cuBLAS | 11.6.5.2 | x86_64, POWER, Arm64 |
CUDA cuFFT | 10.5.2.100 | x86_64, POWER, Arm64 |
CUDA cuFile | 1.0.2.10 | x86_64 |
CUDA cuRAND | 10.2.5.120 | x86_64, POWER, Arm64 |
CUDA cuSOLVER | 11.2.0.120 | x86_64, POWER, Arm64 |
CUDA cuSPARSE | 11.6.0.120 | x86_64, POWER, Arm64 |
CUDA NPP | 11.4.0.110 | x86_64, POWER, Arm64 |
CUDA nvJPEG | 11.5.2.120 | x86_64, POWER, Arm64 |
Nsight Compute | 2021.2.2.1 | x86_64, POWER, Arm64 (CLI only) |
NVTX | 1.21018621 | x86_64, POWER, Arm64 |
Nsight Systems | 2021.3.2.4 | x86_64, POWER, Arm64 (CLI only) |
Nsight Visual Studio Edition (VSE) | 2021.2.1.21205 | x86_64 (Windows) |
nvidia_fs1 | 2.7.53 | x86_64 |
Visual Studio Integration | 11.4.120 | x86_64 (Windows) |
NVIDIA Linux Driver | 470.82.01 | x86_64, POWER, Arm64 |
NVIDIA Windows Driver | 472.50 | x86_64 (Windows) |
CUDA Deep Neural Network (cuDNN)
cuDNN is an NVIDIA library used by deep learning frameworks to accelerate common components, such as pooling, normalization, and backward convolution. Jetpack 5.0 comes with cuDNN 8.3.2, including bugfixes and new enhancements.
TensorRT
TensorRT is a framework created and published by NVIDIA for deep learning inference runtime. Jetpack 5.0 Developer Preview includes TensorRT 8.4.0 Early Access (EA). This EA is for early testing and feedback. For production, the documentation recommends continuing to use TensorRT 8.2.3 or later TensorRT 8.2.x patch. This version of TensorRT has support for CUDA 11.x.
TensorRT 8.4.0 EA has been tested with the following dependencies versions:
- cuDNN 8.3.2
- TensorFlow 1.15.5
- PyTorch 1.9.0
OpenCV
OpenCV is an Open Source library for computer vision tasks. Jetpack 5.0 supports OpenCV version 4.5.4.
Vision Programming Interface (VPI)
VPI is a library provided by NVIDIA to easily take advantage of the power of the Programmable Vision Accelerator and the GPU. It brings Computer Vision and Image Processing algorithms implemented to run the PVA, OFA (Optical Flow Accelerator), GPU, and CPU. Jetpack 5.0 comes with VPI 2.0 version. To learn more about how you can use VPI along with GStreamer, visit our GstVPI section.