Two days ago at Computex Nvidia annunced the new Jetson Xavier, a new embedded board with the robotics as target.
Nvidia Xavier module
It’s been a little more than one year since Nvidia released the amazing Jetson TX2. We were talking about the incredible calculation power available for the first time on an embedded board, about how Computer Vision and Artificial Intelligence task in real time without using a complete Personal Computer were not a dream anymore… after one year they come out with a new embedded board 20 times more powerful and 10 times more efficent!!!
Technical Specs
The new specs are really impressive:
| GPU | 512-core Volta GPU with Tensor Cores |
| DL Accelerator | (2x) NVDLA Engines |
| CPU | 8-core ARMv8.2 64-bit CPU, 8MB L2 + 4MB L3 |
| Memory | 16GB 256-bit LPDDR4x | 137 GB/s |
| Storage | 32GB eMMC 5.1 |
| Vision Accelerator | 7-way VLIW Processor |
| Video Encode | (2x) 4Kp60 | HEVC |
| Video Decode | (2x) 4Kp60 | 12-bit support |
| Camera | 16x CSI-2 Lanes (40 Gbps in D-PHY V1.2 or 109 GBps in CPHY v1.1)
8x SLVS-EC lanes (up to 18.4 Gbps) Up to 16 simultaneous cameras |
| PCIe | 5x PCIe gen4 (16GT/s) controllers | 1×8, 1×4, 1×2, 2×1
Root port and endpoint |
| Mechanical | 100mm x 87mm with 16mm Z-height
(699-pin board-to-board connector) |
The “old” Jetson TX2 was powered by 4 ARM A57 cores plus 2 Denver cores running up to 2 Ghz, the new Jetson Xavier has 8 (!!!) Cortex-A55 ARM v8.2 cores and the incredible amount of 16GB of LPDDR4x RAM.
What makes the Jetson boards unique is the possibility to run massive parallel computation through CUDA and the Nvidia GPU. Everything started with the Jetson TK1 and its 192 Kepler CUDA cores, the first embedded board natively capable of running parallel algorithm based on CUDA, then the Jetson TX1 came with its 256 Maxwell CUDA cores. The Jetson TX2 had again 256 CUDA cores, but the architecture migrates to Pascal.
The new Xavier can rely on 512 (!!!) Volta CUDA cores with support for Tensor Cores and mixed-precision compute that are capable of up to 10 TFLOPS FP16 and 20 TOPS INT8. Impressive!
But that’s not enough, Jetson Xavier is capable of more than 30 TOPS (trillion operations per second) for deep learning and computer vision tasks and has a dedicated image processor, a video processor and a vision processor for accelerating computer vision tasks.
Furthermore it has two NVDLA engines are capable of up to 5 TOPS each, supported by Nvidia TensorRT API for real time deep learning inference. We will be able to use the familiar TensorRT API to leverage the acceleration provided by NVDLA, the same way we did on a GPU.
Software
Nvidia Xavier and Isaac, a new heart for robotics
Looking at software, Jetson Xavier has native support for the new ISAAC SDK for robotic application plus the classical software kit coming with the well known Nvidia Jetpack:
- Linux4Tegra
- CUDA SDK
- CUDNN
- OpenCV4Tegra
- VisionWorks
- TensorRT
- Multimedia API
- and so on…
The dark side…
Unfortunately, there are not only roses and flowers, the new Jetson Xavier module has not the same format of the Jetson TX1 and TX2; Xavier is bigger (about two Jetson TXs put side by side) and the expansion connector is not pin to pin compatible with the older, so we cannot use the same carrier boards.
Furthermore Jetson Xavier has not Wifi and Bluetooth capabilities on board, so the carrier board manufacturers will need to add wireless communication hardware to their new products.
The new Jetson board is surely addressed to professional robotics developers and the price of the developer kit is a very clear clue: $1299 (USD)
Videos from Computex
If you need more information I suggest to you to watch this very detailed video interview:
