“Understanding language is more complex than recognising images,” he said, explaining the choice of a six-core Carmel Arm 64-bit CPU with 6Mbyte L2 + 4MB L3 and 8Gbyte, 128-bit LPDDR4, operating at 51.2Gbps. There is also an Nvidia Volta with 384 CUDA cores and 48 Tensor cores, designed to deliver 21TOPS at 15W or 14TOPS at 10W.
This delivers sufficient compute-performance for running multiple neural networks in parallel while processing data from multiple high resolution sensors.
Equipped with two 4K30 encode and two 4K60 decode channels and supporting up to six CSI cameras and 12 lanes of MIPI CSI-2, the Jetson Xavier NX can be used for training, simulation and inference as part of an end-to-end system using a single system architecture and used to train a robot in a virtual environment and output to a real world application, explained Csongor. In addition to robotics, it can be used for commercial drones, the IoT and industrial automation, production and inspection as well as network video recorders and portable medical devices.
The small form factor and performance parameters allows developers to increase AI capabilities without size or power penalties. The supercomputer is compatible with the Jetson AGX Xavier development kit.
For embedded machines, the supercomputer runs on the same CUDA-X AI software architecture as all of the company’s Jetson offerings and is pin-compatible with the Jetson Nano. It is also supported by the Nvidia JetPack software development kit, which consists of an AI software stack running AI networks and accelerated libraries for deep learning, computer vision and computer graphics.
The Jetson Xavier NX supports TensorFlow, PyTorch, MXNet, Cadffe and other major AI frameworks, says the company. It will be available in March 2020.
