NVIDIA Jetson Orin - SoM Overview
 | NVIDIA Jetson Orin RidgeRun documentation is currently under development. |
|
This section provides an overview of the Orin AGX SoM components and their capabilities.
The NVIDIA Jetson Orin sets a new record on the performance achievable by an embedded system on-edge, delivering up to 275 tera operations per second (TOPS) while keeping power efficiency. The Orin AGX System on Module (SoM) is the small form-factor PCB that encapsulates the computing devices such as the CPUs and the GPUs, the Image Signal Processor (ISP), and the Programmable Vision Accelerator (PVA), the memory, the interfaces for communication and other required system components by NVIDIA that makes this performance possible. Figure 1 shows a picture of the Orin AGX SoM.
Table 1 shows a summary of the hardware characteristics of the SoM:
| Feature | Jetson AGX Orin 32GB | Jetson AGX Orin 64GB |
|---|---|---|
| AI Performance | 200 TOPS (INT8) | 275 TOPS (INT8) |
| GPU | NVIDIA Ampere architecture with 1792 NVIDIA® CUDA® cores and 56 Tensor Cores | NVIDIA Ampere architecture with 2048 NVIDIA® CUDA® cores and 64 Tensor Cores |
| Max GPU Freq | 939 MHz | 1.3 GHz |
| CPU | 8-core Arm® Cortex®-A78AE v8.2 64-bit CPU 2MB L2 + 4MB L3 | 12-core Arm® Cortex®-A78AE v8.2 64-bit CPU 3MB L2 + 6MB L3 |
| CPU Max Freq | 2.2 GHz | 2.2 GHz |
| Memory | 32GB 256-bit LPDDR5 204.8 GB/s | 64GB 256-bit LPDDR5 204.8 GB/s |
| Storage | 64GB eMMC 5.1 | 64GB eMMC 5.1 |
| Deep Learning Accelerator | 2x NVDLA v2.0 | 2x NVDLA v2.0 |
| DLA Max Frequency | 1.4 GHz | 1.6 GHz |
| Vision Accelerator | PVA v2.0 | PVA v2.0 |
| Encoder |
|
|
| Decoder |
|
|
| CSI Camera |
|
|
Graphics Processing Unit (GPU)
The Orin AGX comes with an NVIDIA Ampere architecture GPU. The Ampere GPU brings state-of-the-art graphics and parallel computing techniques, such as ray-tracing, CUDA, and TensorRT support. The Orin AGX 64 GB Ampere GPU is composed of the following modules:
- 2 Graphic Processing Clusters (GPCs)
- 8 Texture Processing Clusters (TPCs)
- 16 Streaming Multiprocessors (SMs), where each has:
- L1 cache of 192KB
- L2 cache of 4MB
- 128 CUDA cores
- 4 Tensor cores (3rd generation)
To learn more about the Jetson Orin AGX GPU, check out our Ampere GPU wiki.
Central Processing Unit (CPU)
Jetson Orin AGX 64 GB comes with 12 Arm Cortex-A78AE CPU cores, each with:
- 64KB instruction L1 cache
- 64KB data cache
- 256KB L2 cache
- Up to 2.2 GHz frequency
The CPU cores are arranged in 3 clusters of 4 CPU cores. Each cluster has a 2MB L3 Cache. The CPU complex also features a 4MB system cache.
Deep Learning Accelerator (DLA)
The Jetson Orin AGX brings NVDLA 2.0, the second generation of NVIDIA's DLA, with nine times the performance of the previous generation. The DLA is designed to accelerate inference in convolutional neural networks, therefore it can be used to free the GPU for other tasks. TensorRT can use the DLA to run inference on various INT8 or FP16 networks.
Programmable Vision Accelerator (PVA)
The PVA engine supports common computer vision kernels: warping, Fast Fourier Transform, image pyramid, filtering, and feature detection. The Orin AGX comes with the second generation PVA v2, composed by:
- 1 Cortex-R5 subsystem
- 2 7-way Very Long Instruction Word (VLIW) vector processing units
- 2 DMA engines
Codecs
Multi-Standard Encoder (NVENC)
Supports:
- H.264
- H.265
- AV1
Multi-Standard Decoder (NVDEC)
Supports:
- H.264
- H.265
- AV1
- VP9
JPEG Codec (NVJPEG)
Supports:
- Color space conversion: RGB to YUV
- YUV420, YUV422H/V, YUV444, YUV400 decoding
Packaging Information
Table 2 shows the packaging characteristics of the Orin AGX SoM.
| Feature | Description |
|---|---|
| Module Size | 100.0 mm × 87.0 mm |
| Connector | 699 pin board to board connector |
| Operating Temperature | Range from -25C to 80C |
| Power Input | 5V to 12V |
| TPP | Integrated Thermal Transfer Plate (TTP) with Heatpipe |
