NVIDIA has unveiled its next-generation edge AI Jetson module and developer kit, the Jetson AGX Orin, promising "server-class" performance in the palm of your hand and a sixfold boost over its predecessor.
"It's been roughly seven years since we first introduced NVIDIA Jetson, and the the traction in the community has been growing leaps and bounds every year," claims Deepu Talla, vice president and general manager for embedded and edge computing at NVIDIA. "So we are we are announcing the next generation Jetson. This is going to be based on Orin SoC. This is going to provide 200 TOPS of AI performance, and the whole device fits in the palm of your hand."
Like previous releases, the Jetson AGX Orin will launch in two flavors. The first is a system-on-module, which is pin- and form factor compatible with the earlier Jetson AGX Xavier — meaning designs built around the older module should be immediately compatible with the new one. The second form is a developer kit, which includes a base board plus module in attractive housing with integrated cooling.
"[Orin is] providing up to six times higher performance in as little as 15 watts for the whole computer — not just the processor, but the whole computer — and up to 50 Watts depending on the depending on how much of the compute power and how much of the different blocks that are necessary to use," Talla continues.
The headline figure of 200 TOPS (trillion operations per second) is, as usual, measured at INT8 precision, and comes courtesy of the Ampere GPU, Tensor Cores, NVIDIA Deep Learning Accelerators (NVDLAs), and vision accelerators. On top of this is a 12-core Arm Cortex-A78AE CPU, the clock speed of which has not yet been publicly announced, and 32GB of LPDDR5 RAM with a 204GB/s throughput rate.
Getting the full performance out of the machine will require pushing that power envelope, however. As is usual for Jetson modules, the Jetson AGX Orin can be configured from 15W to 50W — trading performance for a lower power draw and heat output by slowing down or entirely turning off selected components including CPU cores and GPU hardware.
What NVIDIA has not yet shared is the price, stating only that pricing will be disclosed "when we get closer to availability" — which, all being well, is scheduled for the first quarter of 2022. Interested parties can sign up to be notified about availability on the NVIDIA website.