At the inaugural GPU Technology Conference in Europe, NVIDIA CEO Jen-Hsun Huang unveiled the company’s new Xavier SoC, designed for use in performance demanding self-driving cars.
NVIDIA’s Xavier is the company’s first “supercomputer” chip to be designed using customized ARM CPU cores, instead of an off the shelf Cortex-A part or NVIDIA’s in-house Denver CPU design. The company is keeping the exact details about this news design a secret, but we do know that these cores will appear in a octa-core configuration, alongside 512 CUDA graphics cores from the company’s latest Volta-generation. This 7 billion transistor chip is to be packed using TSMC’s 16nm FinFET manufacturing technology. NVIDIA claims that the processor will deliver 20 TOPS (trillion operations per second) of performance, while consuming only 20 watts of power.
Technical specifications aside, the Xavier is designed with the rapidly expanding and increasingly demanding automotive industry in mind. In particular, self driving cars that require complex and performance demanding deep learning algorithms. This is often dubbed as some form of AI or supercomputing due to the large amount of data crunching involved, something which NVIDIA expects will see a big perfromance boost with its Volta GPU technology.
“Modern supercomputers should be designed as AI supercomputers. This means a system has to be good at computational science and data science and that requires an architecture that is good for both. We want to be able to support models that are very large and process may of those across multiple nodes, so interconnectivity is important. We have shown GPUs that can be shared in this way across massive GPU sets of nodes and the supercomputers of the future will be balanced by these two computational approaches with this architecture.” – Jen-Hsun Huang, Nvidia CEO
Xavier is also designed to meet the ISO 26262 functional safety specification for self-driving cars. There’s also support for 7680×4320 video decoding thanks to the chip’s dual 8K HDR video processors. The company expects the SoC to replace its current high performance DRIVE PX 2 platform. The DRIVE PX 2 is often configured with dual mobile SoCs and dual discrete GPUs, which increases power consumption from 10 to around 20 watts and delivers a similar 24 trillion deep learning operations (TOPS) per second. So Xavier should accomplish similar results with a single SoC solution.
While the Xavier SoC sounds very promising, it will still be a while yet until this hardware appears in a finished product. The chip isn’t even scheduled to start sampling until the fourth quarter of 2017.