NVIDIA recently began shipping a new product, the Jetson AGX Xavier Developer Kit. Looky here:
Jetson AGX Xavier Overview
The Jetson Xavier is the next generation of the Jetson line of development kits. Xavier ups the game from the previous generation Jetson TX2.
The Jetson Xavier introduces a new module format in order to support the increase in bandwidth needs for the next generation of data I/O. This includes introduction of USB 3.1 and PCIe Gen 4.0.
The Jetson AGX Xavier features a NVIDIA Volta 512 core GPU with 64 Tensor cores. Eight Carmel cores (NVIDIA’s own custom 64-bit ARM cores) make up the CPU complex, in 4 dual-core clusters. The cores implement ARMv8.2 with RAS support. All cores are cache coherent, which extends to the GPU and other onboard accelerators.
The Xavier incorporates dual Programmable Vision Accelerators (PVAs) for helping with common computer vision processing tasks.
Another accelerators on the Tegra chip is the Deep Learning Accelerator (DLA), a hardware implementation of the NVIDIA NVDLA architecture.
Some great information about the actual Tegra Xavier chip is available on WikiChip: Tegra Xavier.
The memory subsystem incorporates a 256-bit memory controller which provides high bandwidth LPDDR4 support. 16 GB LPDDR4 Main Memory and 32 GB eMMC Flash memory are integrated on the module.
The Module also supports hardware video encoders and decoders which support 4K ultra-high-definition video at 60 fps in several different formats. Also included is an Audio Processing Engine with full hardware support for multi-channel audio.
Gigabit Ethernet BASE-T is included.
The display controller subsystem allows for three multi-mode (eDP/DP/HDMI) Serial Output Resources. This includes HDMI 2.0a/b (up to 6Gbps), DP 1.2a, and eDP 1.4 (up to 8.1Gbps). There is a HDMI 2.0 connector on the carrier board, the other display ports are accessible via USB 3.1.
Speaking of USB 3.1, there are two USB 3.1 connectors, 40 pins for GPIO, a micro USB connector, a PCIe Gen 4.0 slot. Along with the previously mentioned HDMI 2.0a connector and RJ45 Ethernet port, there is a barrel jack which supplies power to the developer kit. The developer kit can be powered by between 9 and 20V (the kit includes a 65 watt 19V power supply). The developer kit can also be powered through either one of the USB 3.1 connectors.
October 4, 2018 – Dustin Franklin published the slides from a quite wonderful webinar with a lot of hardcore information about the Xavier, including benchmarks. Worth checking out!
Sippy or Speedy
Like the Jetson TX2, there are several different selectable modes for configuring the power consumption and speed. These modes work by managing the number of CPU cores online and setting the frequency of the CPU and GPU cores. The Jetson Xavier has modes similar to the profile of the Jetson TX2 in both low and high performance mode (10W and 15W), along with another Xavier specific super mode (30W). 30W refers to 30 watts, and different variations of modes can tailor the performance in the specific power envelope that you specify.
There are several changes to the Jetson AGX Xavier software stack. The Jetson Xavier runs a Developer Preview of an Ubuntu 18.04 variant named L4T 31.0. The Linux Kernel is 4.9, a newer version than the earlier Jetson TX2 version 4.4. There have been changes to the boot flow. The Jetson Xavier comes with a long list of software libraries, and a good selection of samples with source code.
The new JetPack 4.0 Developer Preview Early Access installer is available to flash and copy system software to the Jetson TX2.
If you are a developer, you can get the Jetson AGX Xavier at a discounted price on the NVIDIA site.
You can buy the Jetson AGX Xavier on Amazon at the regular price. You can also buy the Xavier on the NVIDIA site and other authorized resellers.
Updated January, 2019
Dustin Franklin from NVIDIA has a very nice article about the architecture and performance of the AGX Xavier. NVIDIA Jetson AGX Xavier Delivers 32 TeraOps for New Era of AI in Robotics describes the different functional components of the Xavier including the Deep Learning Accelerator (DLA) and Vision Accelerator. The article also has a comparison between Deep Learning Inferencing Benchmarks between the AGX Xavier and the Jetson TX2.
Here is what NVIDIA is saying about the peak performance of the Xavier:
Jetson AGX Xavier is capable of more than 30 TOPS (trillion operations per second) for deep learning and computer vision tasks. The 512-core Volta GPU with support for Tensor Cores and mixed-precision compute is capable of up to 11 TFLOPS FP16 or 22 TOPS INT8 compute. Jetson AGX Xavier’s dual NVDLA engines are capable of 5 TOPS INT8 or 2.5 TFLOPS FP16 performance each. It also has high-performance eight-core ARM64 CPU, a dedicated image processor, a video processor and a vision processor for accelerating computer vision tasks.
Stay tuned as we begin working with the Xavier to better understand how to take advantage of the next level performance. The amount of documentation available now is impressive, you can find links on the Jetson AGX Xavier forum.