You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Of course we all knew this when Jim Keller joined Tesla In January 2016 as Vice President of Autopilot Hardware Engineering. Jim was at Apple through the PA Semi acquisition where he was part of the A4 and A5 team. And we can see Tesla on SemiWiki analytics shortly thereafter. CNBC sources say Elon Musk confirmed Tesla (NASDAQ:TSLA) is working on its own AI hardware to reduce dependence on Nvidia (NASDAQ:NVDA).
Musk made the comments during a talk at a company party last night. His guest was Telsa’s VP of hardware Jim Keller, who was a chip architect at AMD. Tesla’s current Autopilot hardware system utilizes Nvidia graphics cards. Musk and Keller cited power and cost control and better efficiency among the benefits of using custom hardware.
All the best to Tesla in trying to create their own CPU, GPU or TPU to compete with commercial offerings already in the marketplace for autonomous vehicles. Tesla would have to engineer a unique architecture to make it worth the investment, then we have to wait at least two years for them to build a team of SoC experts, tape out, debug, program and then announce to the world their new chip(s). With so many semiconductor companies trying to serve this market segment already, can Tesla really create yet another new hardware, software and OS system to create their own platform?
Well, someone must think they have a decent idea for a new GPU/TPU/CPU for Telsa to want to jump into this market. Perhaps there is some algorithmic approach for object recognition that it is not being served yet.
I doubt Tesla is making a GPU. My guess would be an SoC specifically for ADAS. Right now there are dozens of microprocessors independently controlling different pieces of the car's operations. Tesla seems to want to have a master SoC for easy upgrades and such.
There's talk about the performance requirements for reacting to 3D data scans in real time (0.1 s or better) for higher levels of automated driving. They are out of this world. Hundreds if not thousands of independent, deterministic logic cores.
It's a pure, high performance logic scenario, not GPU-compute.
Tesla likes to vertically integrate and also likes to solve big problems by themselves. Both vectors converge on this new chip.
I think you've hit it on the head here. I've never thought GPUs were the best solution for autonomous driving, just the only available solution at the time. Just because an architecture is good at turning thousands of software objects into a 3D image doesn't mean it is good at turning a 3D image into discrete objects. I'm sure a better architecture will appear in due course. Indeed it has to because the power drain of current solutions, even when scaled to 5 or 3nm, is going to restrict usage to the larger petrol/diesel cars with large alternators that we're supposed to be trying to get away from
