Question from a noob.
What hardware FSD needs to perform? To gather data, HW3 or 4 with 8 cameras, etc.
But to make the car go? the same hardware? does it need cameras?
Could Tesla really license FSD to different cars as of now?
So.. I'll try and make this short.
The hardware that has the smarts to do FSD in Teslas is something truly new in the world: A built from first-principles pair of neural networking devices, working in tandem, below the dash. These were designed from scratch
by Tesla. The computer in the Tesla is water-cooled and dissipates a fair amount of heat. Just so we're clear on this: A neural network, like the one inside your skull, works by combining weights of a large but variable number of inputs, then generating outputs that can go into more neural networks; feedback is often involved. What these things are
not are digital computers, like the laptop, cell phone, or PC that you may be using. Like many things in engineering, neural networks can be very, very good at solving certain problems and not so good at solving other problems. The same can be said for digital computers: If you want to add and subtract numbers, digital computers are your friends and can do those functions at an astonishing rate (like, billions of times a second).
It's possible to take a neural network and make it add and subtract numbers; it's possible to take a digital computer and have it compute weights.
As it happens, the nearest thing to an honest-to-golly supercomputer that can be bought for less than a zillion dollars turns out to be graphics cards. The original purpose of these things was to provide computer game players realistic looking graphics. When Tesla first got started, they got themselves graphics card chips and did neural networking for images processing and such. They then realized this wasn't going to fly for the purposes of driving around the landscape, got themselves a ridiculously bright bunch of scientists and IC designers (one, notably, from Apple), and designed the first version of the current hardware.
I remember that the video chip company that Tesla abandoned was rather miffed since the chip company had designs on this whole self-driving computer business.. but their devices, no matter how speedy, or how well their marketing was working, was four or five times slower than what Tesla came up with, if memory serves.
Now, all that foo-fer-aw was back in 2018/2019. Tesla has continued to improve their devices. Other manufacturers of ICs for the trade has, by now I'm sure, put some basic neural networking functions on silicon. So, it's not like Tesla doesn't have competition these days. On the other hand: They have a
lead.
So: A lot of image processing. A lot of neural networking with integrated digital computers. It's amazing that it's cheap enough to put all that in a car.
And then.. there's the software. Yes, there are the likes of Google, Apple, GM, VW, and lots of other folks trying to build self-driving cars. Tesla is still out in front of all of these. Remember: They're running some ridiculously complicated code on top of computers
of their own design, the HW being optimized to run the SW at $DIETY's own speed.
So.. This is not Windows or iOS that one can develop a simple app upon and drop it into play and expect it to run everywhere. My guess is that, if Ford wanted to use Tesla's software, they'd likely have to use Tesla's hardware as well.
Don't know if you're aware: But those robots walking around in Tesla's development labs? They're running the same computer that's in a Model 3. Because the computer, with the right software, is
that capable. Fun.
