Tesla Self-Driving Chip: stripped-down version of Google TPU?

Google
anononly

Go to company page Google

anononly
Apr 23, 2019 50 Comments

Elon claimed it’s 21x faster on achievable performance metric compared to Nvidia, which I feel kinda stretch, but doable. Peak OPS/FLOPS doesn’t matter without sustainable utilization a chip can deliver for real workload.

Is this what Jim Keller was hired for? As he left to work for Intel last year (likely to pioneer MobileEye), I guess we might seeing TPU variants in future Intel/MobileEye chips? How’d that fare compared to Nervana accelerator?

@nvidia @tesla

comments

Want to comment? LOG IN or SIGN UP
TOP 50 Comments
  • Apple
    metafor

    Go to company page Apple

    metafor
    It sounds like it’s a more special purpose processor. Only int8 is supported and it only has 6 instructions. No memory coherency (all handled by software) and buffering/dataflow is specifically tuned for the 8 cameras + ultrasonic + radar bandwidth of a Tesla. Plus batch size of 1 frame.

    Compared to TPU which accelerates both learning and inference and has to support programming from a framework like TensorFlow. And is meant for batch jobs rather than real-time frame by frame jobs.
    Apr 23, 2019 7
    • Google
      Woopsy

      Go to company page Google

      Woopsy
      It does not have much more than 6, really.
      And yes it is likely closer to TPUv1 if you want to compare
      Apr 24, 2019
    • Intel
      Hunting

      Go to company page Intel

      Hunting
      fp is not needed for object recognition I am guessing...you need parallel processing for simpler data, than single processing for complex ones.
      All guesses though.
      Apr 25, 2019
  • It's not accurate to call Tesla's chip a stripped down version of Google's TPU, the same way it wouldn't be accurate to call AMD's cpu a stripped down version of Intel's cpu.
    Apr 23, 2019 6
  • Nvidia
    bdei51

    Go to company page Nvidia

    bdei51
    It's an inference processor with a power limitation derived from car's requirements, it is not a stripped down version of anything. TPU is a data center chip that supports both inference and training (v2 and v3 at least). Quite different.

    And yeah, don't believe the numbers Elon quotes. They are wrong, and Nvidia published a blog post related to Elons claims.
    Apr 24, 2019 7
    • Apple
      metafor

      Go to company page Apple

      metafor
      Another thing is that they obviously benchmark at how many frames/sec of the specific frame sizes of a Tesla using int8 inference when using what looks like a basic classification neural net.

      While the Xavier likely has a theoretical peak of 30 TOPS, Tesla is probably quoting what they managed to achieve using their own workload.

      Notice that a good deal of time was spent during the presentation talking about the data flow and on-chip SRAM. And how much it was emphasized that this was a very single purpose chip without many modes or computational flows. It would not surprise me if they were able to hit near peak theoretical TOPS with their own chip.
      Apr 26, 2019
    • Seagate
      iamwilliam

      Go to company page Seagate

      iamwilliam
      The blog only discusses performance number. Of course it is possible to tune up the performance by increasing the voltage. On the flip side, it will be a hot chip.
      Apr 27, 2019
  • Elon is always lying. This FSD promo is just to paint the news cycle bc Wednesday their numbers are going to be horrible. He can’t build a profitable car company so he wants to be valued on autonomous driving lol.

    Elon doesn’t care about this company. That’s why he embezzled from it to bailout his and his cousins insolvent solar city company. It about money. Elon has taken margin loans against his Tesla equity and cannot have the stock price go down too much, even if that’d be good for the company.

    https://twitter.com/richardkeppler1/status/1120888716132790272?s=21

    3/19/15 (NYT) -"Musk Says Self-Driving $TSLA Cars Will Be in the U.S. by Summer"

    10/19/16 (Wired) -"MUSK SAYS EVERY NEW TESLA CAN DRIVE ITSELF"

    1/24/17 (Electrek) -"Tesla to transition from ‘Enhanced Autopilot’ to ‘Fully Self-Driving’ as soon as ‘3 to 6 months’, says Musk"
    Apr 23, 2019 2
  • Pure Storage / Eng
    xluy52

    Go to company page Pure Storage Eng

    xluy52
    Bro, it’s an ASIC. TPUs can handle any model. The ASIC only works on one model and can optimize for that
    Apr 24, 2019 4
    • Pure Storage / Eng
      xluy52

      Go to company page Pure Storage Eng

      xluy52
      @cisco, right, but it’s an ASIC that has overhead caked in to be able to generalize to any tensor graph. You can yield a lot more speed out of a chip if it will only ever have to do the same computation on the same type of data.
      Apr 25, 2019
    • Nvidia
      bdei51

      Go to company page Nvidia

      bdei51
      Much of that overhead is on the compiler I would say
      Apr 25, 2019