> they could have used off-the-shelf components for the same task.
They did use off-the-shelf components. They performed poorly and they resulted to developing a custom die with the right components to handle the task of computer vision. They explicitly stated this in multiple interviews. The off-the-shelf solution was crap.
AI/ML components change rapidly. What happens if they decide to use an entirely different approach. They develop a new chip? Still seems excessive to develop custom chips for unfinished/experimental tech like self-driving: https://www.theverge.com/2019/4/24/18514308/tesla-full-self-...
Point is they are doing it. Tesla does it. Apple does it. Microsoft does it. They are all developing their on hardware, and it's paying off, especially in Apple's and Tesla's use case. Research and development is the cornerstone of new innovation. Who cares if it's new tech. Not sure what you mean by "unfinished". They have a product that works. It's new and could use improvements, but Point is, they are doing it and winning. We can't sit around and wait for Ford and GMC, but these guys are being ran by dinosaurs. Stuck in the stone age, waiting for ICE to make a comeback. Here's the crutch of it all, ICE is pretty much out, with the old waterfall model, dead, old tech, EV is far more simple and a heck of a lot better. This whole ICE mentality where you release a car and it's "complete", has to go. This is why we have 'recalls'. Your car should continue to receive updates and improve as you drive it off the lot. Security updates are necessary, UI updates are necessary, performance and economy updates are necessary.
> This is why we have 'recalls'. Your car should continue to receive updates and improve as you drive it off the lot. Security updates are necessary, UI updates are necessary, performance and economy updates are necessary.
Did you ever try to rollout change to millions of units of anything, with individual owners of each item, operating in vastly different environments and with additional modifications from those owners?
Rolling updates is A HUGE issue, if you need reliability. It’s not a random app. It’s, for majority of the people, second most expensive purchase in their live (only behind a house). Messing up with that is very very risky.
This is less of a risk than it might appear. To be sure, it would be expensive to replace the computer part of the entire installed fleet. But the (potentially) full self-driving option is very richly priced, and sells with an incremental margin of 100% when component replacements are excluded. The sensor & computer package are included with all production cars, regardless of whether the FSD option is purchased.
Additionally, Tesla still grows their sales and customer base exponentially. This means that most significant changes in strategy during the production ramp-up will require (in the form of deprecation of now-obsolete production resources) only a small fraction of the investments required to achieve their full, steady-state production capacity.
The same argument holds for the potential case of major changes in battery technology. (Replacements of current fleet excluded, which would not happen for batteries). If Tesla's current battery technology is obsolete in 5-10 years, that's not a very big deal as their sales are expected to be a multiple of today's by then, and the new technology would be phased in during the expected production ramp. (If Tesla fails to grow by a multiple of today's sales, they have largely failed).
Based on their AI chip presentation, the important thing is that they have a software/hardware integrated design relationship. Your comment assumes they do this once and never again.
I seriously doubt Tesla is not actively improving the chip, and if they need a fundamentally new software approach, then they have the process/iteration loop established.
They didn't. They had been saying for years that it would require a computer upgrade that could be done with a service visit, and that the service visit would be free for people who purchased the "FSD" option.
They absolutely did, and then changed the story when they announced they were building HW3.
> All Tesla Cars Being Produced Now Have Full Self-Driving Hardware
> Oct 19, 2016
> To make sense of all of this data, a new onboard computer with more than 40 times the computing power of the previous generation runs the new Tesla-developed neural net for vision, sonar and radar processing software
^ this refers to the nvidia thing 'onepremise says was too bad for the job.
I don't have time to dig it all up, but they were talking about the potential for computer upgrades as far back as 2016 as well: https://twitter.com/elonmusk/status/789008557341454336 (though not for pre-2016 vehicles).
I can see where people who don't follow the company closely wouldn't have been aware of this, though. And, obviously there have been plenty of false and misleading statements made in this area.
Power and capability. Custom silicon means they can build the chip with all of the chip's surface being used for things their software needs (ie. extra surface for parallel computing and zero surface for unneeded features). They can also tune the chip with ops/watt in mind. Nvidia has touted their A.I. chips being x10 more powerful, but they also took twice the power to do it [1]. For an electric vehicle that lives on only the charge stored in its batteries, you want to use exactly as much power as you need to get the job done and no more.
They did use off-the-shelf components. They performed poorly and they resulted to developing a custom die with the right components to handle the task of computer vision. They explicitly stated this in multiple interviews. The off-the-shelf solution was crap.