Nvidia presents improved Xavier SoC for autonomous machines
Besides professional and gaming graphics solutions, Nvidia has recently taken an interest in artificial intelligence and neural network applications that expand the way autonomous machines interact with the environment. Derived from las year’s Volta architecture, just like the new Turing architecture featured in the freshly launched gaming GPUs, the new Xavier system-on-chip for autonomous machines introduces the Carmel architecture that powers the processing unit of the chip.
The latest Xavier SoC was presented at the Hot Chips summit as using the 12 nm FFN manufacturing process from TSMC and coming with numerous specialized processors such as:
• Carmel CPU with eight custom ARM 8.2 cores
• Volta GPU with 512 CUDA Tensor Cores
• Deep Learning Accelerator capable of 11.4 deep learning int8 TOPs
• Programmable Vision Accelerator capable of processing 1.7 TOPs
• Image signal processor with native HDR capabilities and 2.4 GPixel/s processing
• Multimedia accelerators (stereo sound, optical flow)
• Lens distortion correction hardware
• Support for 256-bit LPDDR4x/5
• 20 GB/s NVLink with multi-PCIe Gen 4.0 controllers
• Support for 4x displays with HBR3 and HDMI 2.0 outputs
• a total of 9 billion transistors over a surface of 350 mm2
Regarding scalability, the Xavier SoC can power single chip variants (Jetson Xavier), as well as multi-chip versions including the Drive Xavier and Drive Pegasus versions. Nvidia is already working with Tesla and AI industry leaders to test the Xavier and improve the architecture. Additionally, the new SoC is designed to be energy efficient, but Nvidia did not provide any TDP figures for the time being.