NVIDIA Announces Compact Jetson Xavier NX "AI Supercomputer"

NVIDIA has announced the latest entry in its Jetson high-performance artificial intelligence family, and what it proclaims to be the "world's smallest supercomputer for AI at the edge:" the Jetson Xavier NX.

"AI has become the enabling technology for modern robotics and embedded devices that will transform industries," claims Deepu Talla, vice president and general manager for Edge Computing at NVIDIA. "Many of these devices, based on small form factors and lower power, were constrained from adding more AI features. Jetson Xavier NX lets our customers and partners dramatically increase AI capabilities without increasing the size or power consumption of the device."

As the name implies, the Jetson Xavier NX is based on the company's previous Jetson AGX Xavier computer-on-module (COM). Where the new Xavier NX differs is in size and power draw: The NX measures just 70x45mm (around 2.76x1.77") and can be configured to operate at a 10W or 15W power envelope.

Less power, of course, means lowered performance: The full-size Jetson AGX Xavier boasts 32 trillion operations per second (teraOPS) of INT8 compute, contributed by the eight-core 64-bit NVIDIA Carmel ARMv8.2 CPU, 512-core Volta GPU with 64 AI-accelerating Tensor Cores, and two Deep Learning Accelerators (DLAs), all of which can make use of the 16GB of on-board RAM; the Xavier NX, by contrast, offers 21 teraOPS in its 15W configuration dropping to 14 teraOPS in 10W mode, provided by a six-core Carmel CPU and 384-core Volta GPU with 48 Tensor Cores, though the design retains the two DLAs. The Xavier NX cuts the maximum available memory in half to 8GB, however, while losing the seven-way very long instruction word (VLIW) vision processors of its bigger sibling.

Lower performance doesn't mean poor performance, however: NVIDIA claims that the Jetson Xavier NX comes a very close second to the full-fat Xavier in a range of benchmarks, while offering pin compatibility with the company's low-cost maker-centric Jetson Nano COM. The design also benefits from full compatibility with the rest of the Jetson family, meaning that any project can be developed on any Jetson model — resources permitting — and then transferred to any other model to improve performance or lower cost.

NVIDIA's latest Jetson also includes 16GB of on-board eMMC storage, 12 MIPI CS-2 lanes configurable as 3x4 or 6x2 for six cameras and up to 36 virtual channels, gigabit Ethernet, four PCI Express Gen. 3 lanes with up to 80 gigatransfers per second (GT/s) of bandwidth total, one USB 3.0 10Gb/s and three USB 2.0 ports, and one SDIO, two SPI, three UART, two I2S, four I2C, one CAN, and general-purpose input/output (GPIO) on-board, all accessed through a 260-pin edge connector.

The Jetson Xavier NX is due to launch in March 2020 priced at $399 in volume, NVIDIA has confirmed, with per-unit pricing yet to be confirmed; that's a solid $200 less than the larger AGX Xavier's 8GB variant though $150 more than the earlier Jetson TX2 family's entry point.