people forget tesla previously used nvidia chips for inference compute on its vehicles (nvidia drive px2) - this was "hardware 2" tesla then said nah, we can do something better. they designed (manufactured by samsung on 14nm) hw3 and it was in cars in 2018. custom designed > off the shelf six years later, that same hw tesla started designing in 2016 is running the most powerful autonomous driving platform on earth. the entire suite (hw3 and 8 cameras) probably costs ~$1,000 as @jamesdouma recently explained, v12 is much more optimized for hw3 than v11 was (serendipitous for tesla). hw3 has 3 core components: neural net processing units which are ~100x faster than the gpu's which are themselves ~100x faster than the cpu's the 300k+ lines of c++ code were primarily running on the cpu's (the weakest link) - now with most of that code gone, the video data is primarily running on the nnpu's and gpu's - the more capable parts of the chip tesla's hw3 also consumes less than 100w of power. competitor data is opaque (nothing to brag about) but it's likely multiples higher and thousands of $ more expensive oh and $tsla already has hw4 ready to be called upon in production vehicles which is even more capable legacy oem's can either waste billions trying to catch a moving target or they can license by far the best autonomous suite out there elon has hinted tesla would license the tech for free. ford outfits its cars with 8 cameras and hw3 (or 4) and charges its customers the same price for fsd the reason this model can work is the low cost of tesla's system. as long as tesla takes a cut greater than ~10% (they will) it would be a profitable move. plus tesla maintains access to and control of the data ford obviously is a big winner in this scenario as well (ends with a profit from selling fsd even after giving up a cut to tesla and accounting for licensing fees to tesla) by the time ford (or any company) gets this system built into production the system may be very close to robotaxi level if not fully there we're about to find out which oem is serious about autonomy and survival in the next 12 months
@DillonLoomis22 I'm waiting for Ford to back out of their Red Bull F1 engine deal. Until that happens, I remain skeptical about how serious they are about going fully electric and autonomous. What does Ford gain by spending billions on racing an outdated technology?
@DillonLoomis22 Nice from Jensen / NVDIA to deliver servers with super computer speed to Tesla after Elon asked for help to solve FSD finally. Dojo could not do it at this point ( and might never ).
@DillonLoomis22 Agree that Ford is likely, especially w/ their just announced EV pause. I do wonder about liability. Modern supplier contracts specify quality/recall cost responsibility. Negotiating FSD with a competitor based on somewhat unknown liability laws could become a huge sticking pt
@DillonLoomis22 That's right GPUs aren't specialized. Tesla silicone is specialized and is low power. That's critical for EV range. The fact that it's cheap is only upside for Tesla. Recommended listening to the @jamesdouma interview 🙏🙏🙏
@DillonLoomis22 People just love to forget important stuff in general when it comes to Tesla.
“8 cameras & the Hardware” is a massive over simplification . For a Hardware platform to be an open Bi-directional Connected Car that is also S/W defined, Hundreds of components need to be single sourced or home grown to run the Tesla Operating System as the foundation layer onto which FSD (+’+ +) can be loaded. It’s a Very Big SKUNK.