Nvidia AGX Xavier puts the brains in autonomous robots
A lot of talk around here revolves around powerful parts suitable for your latest personal computer, phone, workstation, or server. That list doesn't cover all the use cases for a juicy, tasty chunk of processing power. Enter the latest member in the company's Jetson family of "system-on-modules," the Nvidia Jetson AGX Xavier.
Nvidia says the AGX Xavier board has 10-W, 15-W and 30-W configurable operating modes, and is fit for robots, computer vision, medical instruments, and autonomous machines, among other applications. The company says Xavier offers the performance of a workstation in a board measuring 3.9" by 3.4" (10 x 8.7 cm). That claim is certainly bold, but then again, the system's specs are also generously sized.
According to Nvidia, Xavier's neural-network output is 32 TOPS. That impressive figure comes by way of Xavier's centerpiece, a Volta GPU with 512 shader ALUs and 64 tensor cores. The cerebellum in the chip is an eight-core Nvidia Carmel ARM v8.2 processor with 8 MB of L2 cache and 4MB of L3. The CPU is connected to 16 GB of LPDDR4x on a 256-bit bus.
The 9-billion-transistor Xavier includes two Deep Learning Accelerator engines and a vision processor. The module's video-handling capabilities can deal with up to four streams of HEVC 4K video at 60 FPS. Additionally, the SoC offers 16 lanes of PCIe 4.0 connectivity along with a bevy of miscellaneous I/O including Ethernet, multiple display outputs, USB, and an alphabet soup's worth of peripheral connectivity acronyms.
It's not difficult to think up more than a few use cases for this kind of tech, and Nvidia's PR notes an interesting one. Oxford Nanopore, a startup focusing on DNA sequencing, apparently built a hand-held computer based on the AGX Xavier. The startup named the device the MinIT, and claims that it can run DNA sequencing up to 10 times faster than a "standard laptop."
Interested developers can order Jetson AGX Xavier boards for the amount for $1099 a piece, so long as they're buying at least 1000. The price may be steep, but we figure the performance available on tap likely justifies it.