Nvidia’s 70 x 45mm “Jetson Xavier NX” module runs Ubuntu on a hexa-core Arm SoC with a 384-core Volta GPU and delivers 14 TOPS (10W) or 21 TOPS (15W) AI performance. By comparison, the larger, octa-core Xavier AGX has 512 Volta cores and up to 30-TOPS AI.

To bolster its edge AI portfolio, Nvidia unveiled a stripped down version of its high-end, 105 x 87 x 16mm Jetson Xavier AGX module. Due to ship in Mar. 2020 for $399, the Jetson Xavier NX is more powerful than the 87 x 50mm Jetson TX2 in both its CPU and GPU while shrinking to the 70 x 45mm size of the lower-end Jetson Nano.

Jetson Xavier NX
(click image to enlarge)

The Jetson Xavier NX cannot match the low power consumption of the Nano, which can run on as little as 5 Watts, but it can operate at 10W with up to 14 TOPS (trillions of operations per second) of AI performance generated from the CUDA-X AI libraries running on its Volta GPU.

The Xavier NX can deliver 21 TOPS at 15W, which enables it to “run multiple neural networks in parallel and processing data from multiple high-resolution sensors simultaneously,” says Nvidia. By comparison, Intel’s Movidius Myriad X AI accelerator has 1 TOPS and Arm’s Ethos-N77 NPUs goes to 4 TOPS, but both support lower power operation.

Here is a quick comparison of the Jetson lineup, ranked in order of ascending processing power:

• Jetson Nano — 69.6 x 45mm; 4x -A57 @ 1.43GHz CPU; 128-core Maxwell GPU; 4GB LPDDR4; 16GB eMMC
• Jetson TX2 — 87 x 55mm; 2x Denver and 4x -A57 CPU; 256-core Pascal GPU; 8GB LPDDR4; 32GB eMMC; opt. WiFi/BT
• Jetson Xavier NX — 69.6 x 45mm; 6x ARMv8.2 Carmel CPU; 384-core Volta GPU; 48 tensor cores; 8GB LPDDR4; 16GB eMMC
• Jetson Xavier AGX — 105 x 87mm; 8x ARMv8.2 Carmel (-A75 like); 512-core Volta GPU; 64 tensor cores; 16GB LPDDR4; 32GB eMMC

Applications include “embedded edge computing devices that demand increased performance but are constrained by size, weight, power budgets or cost,” says Nvidia. These include small commercial robots, drones, intelligent high-resolution sensors for factory logistics and production lines, optical inspection, network video recorders, portable medical devices, and other industrial IoT systems.

The Jetson Xavier NX has twice the RAM of the Nano with 8GB, and similarly supplies 16GB eMMC. It’s pin-compatible with the Nano, enabling Nano-based carrier boards and other hardware to upgrade.

The Xavier NX is equipped with 6x ARMv8.2 “Carmel” CPU cores with 6MB L2 and 4MB L3 cache compared to 8x ARMv8.2 Carmel cores with 8MB L2 and 4MB L3 cache on the Xavier AGX. (AnandTech has benchmarked the Xavier AGX’s CPU and found it to be roughly equivalent to a Cortex-A75, compared to Cortex-A57 on the quad-core Nano.) No clock rates were mentioned, and it’s possible these are throttled down Carmel cores optimized to run at lower power.

Jetson Xavier AGX (left) and Jetson Nano
(click images to enlarge)

Like the Xavier AGX, the Xavier NX has a high-end Volta GPU for running CUDA-X AI algorithms and advanced graphics routines. However, there are only 384 cores compared to 512. The module similarly provides 2x NVDLA deep learning engines., but it is limited to 48 tensor cores instead of 64 on the AGX, There’s no mention of the AGX’s 7-way VLIW vision accelerator co-processor.

Nevertheless, the GPU appears to be considerably more powerful than the TX2’s 256-core Pascal graphics and the Nano’s 128 Maxwell cores. Like the Xavier AGX, the NX supports dual 4Kp60 video decoding, or up to 32x [email protected].

The Xavier NX is limited to dual 4Kp30 encoding instead of the AGX’s dual 4Kp60. The module supports up to 6x MIPI-CSI cameras (or 36 via virtual channels) at 1080p60 resolution and 12 lanes (3×4 or 6×2) of MIPI-CSI-2 at up to 30Gbps. By comparison the AGX supports 16 lanes of CSI-2 and 8x lanes of SLVS-EC for up to 16 simultaneous cameras.

The module ships with 8GB 128-bit LPDDR4x at 51.2GB/second compared to 16GB 256-bit LPDDR4 on the AGX. On-module storage is similarly halved to 16GB eMMC 5.1 down from 32GB.

Nvidia Jetson Xavier Developer Kit

The Xavier NX has a GbE controller, but there’s no mention of onboard WiFi/BT, which is found on the Jetson TX2, but not the Nano. There support for dual multi-mode 4K display interfaces with support for DP 1.4, eDP 1.4, and HDMI 2.0.

The only other interfaces expressed via the 260-pin SODIMM connector that are listed on the spec list are PCIe Gen 3 x1 and x4 with “root port and endpoint” support. However, the announcement also mentions the presence of “low-speed I2Cs and GPIOs.”

A developer kit appears to be in the works. In the meantime you can apply a patch to the Jetson AGX Xavier Developer Kit that emulates Xavier NX performance. The module runs on an Ubuntu based Linux distribution and is supported by the Nvidia JetPack SDK, which includes an AI software stack. Like other Jetson modules, the Xavier NX supports CUDA-X AI plus AI frameworks including TensorFlow, PyTorch, MxNet, and Caffe.

 
Further information

The Nvidia Jetson Xavier NX module will be available in Mar. 2020, priced at $399. More information may be found in Nvidia’s announcement and product page.