4 votes

Telsa unveils internally developed neural network accelerator chip and drop in replacement for existing Nvidia solution

1 comment

  1. Neverland
    (edited )
    Link
    UPDATE: Doh, it did come out in the press while I was listening to the call: https://electrek.co/2018/08/01/tesla-chip-most-advanced-computer-autonomous-driving-autopilot-hardware-3-update/...

    UPDATE: Doh, it did come out in the press while I was listening to the call:
    https://electrek.co/2018/08/01/tesla-chip-most-advanced-computer-autonomous-driving-autopilot-hardware-3-update/

    https://www.cnbc.com/2018/08/01/elon-musk-says-tesla-ai-chip-project-is-finally-coming-to-fruition.html


    I have not seen this in the press yet, so here is a link to the part of Tesla's earnings call today which I thought you all might find interesting. Pete Bannon, who worked at DEC, Intel, Apple (led design of iPhone ARM chips), has designed the Tesla Hardware 3 platform which is an order of magnitude improvement over the NVIDIA GPU solution. In the linked part of the video they describe the solution and why their bottom-up neural network chip design is a 10x processing improvement over the leading GPU solution.

    I remember reading that someone from NVIDIA said that they did not think that Tesla could deliver on their self-driving ambitions with NVIDIA's platform, well... they couldn't and so they just secretly built their own optimized hardware.

    Disclaimer: I hold no positions on TSLA stock. Sorry about the title gore, wish I could edit it.

    edit: This must be a Tesla Autopilot specific ASIC, correct? Along the lines of Google's Tensor Processing Unit?