When people think about NVIDIA, they think either AI/ML/DL, automotive or gaming. In fact, Moor Insights & Strategy ML analyst, Karl Freund, wrote this week how “NVIDIA ran the tables on MLPerf”, an industry-leading ML benchmark. NVIDIA is also a player in the robotics and embedded AI space as well and last week NVIDIA made news with what it touted as “the world’s smallest, most powerful AI supercomputer,” the Jetson Xavier NX, designed expressly for robotic and embedded edge computing devices in a 10-15W envelope.
Enabling edge computing
AI computing at the edge presents a very specific set of chip requirements—many of these devices have increasingly high-performance requirements, with weight, size, power and cost constraints. Jetson Xavier NX seeks to enable such devices at 10-15W platform power draw, which include the likes of small commercial robots, drones, optical inspection, network video recorders, portable medical devices and sensors for factory logistics, production lines, and other various industrial IoT applications.
First off, Jetson Xavier NX is tiny—at 70×45 millimeters, it is smaller than the size of a credit card. Second, it packs quite a punch—as much as 14 TOPs, at 10 watts, or 21 TOPS, at 15 watts. For reference, NVIDIA says it can achieve as much as 15 times the performance of the Jetson TX2, yet in a smaller size, using the same power. In my experience, any time one sees this large of a jump in performance in the same form factor and power, great things happen.
Xavier NX is capable of running multiple neural networks in parallel and can process data from multiple high-res sensors simultaneously. This is important, for example, if you had a robot with multiple cameras, gyroscopes and even the ability to smell.
Jetson Xavier NX consists of an NVIDIA Volta GPU, with 384 CUDA cores, 48 tensor cores, and 2x NVDLA. The increased number of GPU cores should enable the device to perform training as well as inference, a valuable feature that significantly expands its utility. Sure, it’s lighter training than one would see in a car or datacenter, but enough to improve cognition in near real-time as the environment changes without having to phone home for an inference model or upload data.
Its CPU is a 6-core Carmel ARM 64-bit CPU, and for memory, it features 8GB 128-bit LPDDR4x, at 51.2GB per second. As far as its video capabilities go, it features 2x 4K30 Encode and 2x 4K60 Decode capabilities and supports as many as 6 CSI cameras (36 via virtual channels), as well as 12 lanes MIPI CSI-2 (3×4 or 6×2). It features Gigabit Ethernet connectivity, and supports Ubuntu-based Linux OS. NVIDIA touts, which is fair, Jetson Xavier NX’s rich set of IOs including the previously mentioned CSI camera inputs, PCIe, USB and low-speed I2Cs and GPIOs. All of this should make Jetson Xavier NX compatible with many different peripherals and sensors, adding to the supercomputer’s draw.
AI embedded computers are software-defined machines
One of the keys to NVIDIA’s success is its software platforms. It invested huge and early into the basic ML plumbing and upped the game with specific vertical libraries which I believe has kept a lot of the competition scrambling. To aid in adoption and make it easier for companies who are already in the embedded market, Xavier NX runs on the same software architecture as its other Jetson products, CUDA-X AI, and runs all CUDA-AX AI models. As far as software goes, it includes NVIDIA’s JetPack SDK. As described by NVIDIA, JetPack is complete AI software stack capable of running complex AI networks, computer graphics, computer vision, multimedia, accelerated libraries for deep learning and more. I see this software stack, and particularly the compatibility it brings, as a major differentiator from other AI accelerators on the market.
The Jetson Xavier NX joins the rest of NVIDIA’s powerful Jetson family, including the Jetson AGX Xavier series (read more here), the Jetson TX2 series (read more here), and the $99 Jetson Nano, with which the Jetson Xavier NX is pin-compatible. It’ll be available this coming March for $399, but developers can get a jump start on app development using the Jetson AGX Xavier dev kit.
All in all, the Jetson Xavier NX looks to be a great addition to NVIDIA’s Jetson line, and with its small form factor, low power, high performance and mature tool chains, I think it will be an appealing product for engineers playing in the edge embedded computing and robotics realms. Once again, it looks like NVIDIA’s leading the pack in this power, size and performance space, especially when one factors in software.