AI making AI better.
The training of the software matters a lot, but AI making better AI hardware is an indication that this stuff can bootstrap itself perhaps better than we can do the job.
Samsung is building a huge fab - maybe $16 billion? - in Texas for these chips. They’re even thinking of using the compute available remotely. Think of all those Tesla cars sitting around on chargers.
I’d prefer not to have spyware in my car
I am a computer hobbyist and enjoy playing with AI models known as LLMs (large language models) and using various AI apps that do image, video and music generation, along with chat.
AS an example here is my humorous cat music page which uses AI generated images and music just for fun.
At the end of page 2 there are also a picture of me with my Skypup ultralight airplane along with a couple pictures that I took that day from it. This was decades ago before drones as we know them these days were available.
The amount of nonsensical hype about AI is somewhere above 95% when it comes to anything related to it.
If Musk and Intel team up it could be done with ease. Both companies are American - in culture and production facilities.
If I were to bet my life savings on who the winner of the AI development would be, I would bet on Musk.
HW3 is based on a custom Tesla-designed system on a chip called “FSD Chip”, fabricated using a 14 nm process by Samsung. Jim Keller and Pete Bannon, among other architects, have led the project since February 2016. FSD Chip features twelve ARM Cortex-A72 CPUs operating at 2.6 GHz, two systolic arrays (not unlike the approach of TPU) operating at 2 GHz and a Mali GPU operating at 1 GHz. Tesla claimed that FSD Chip processes images at 2,300 frames per second (fps), which is a 21× improvement over the 110 fps image processing capability of HW2.5. The firm described FSD Chip as a “neural network accelerator” custom-designed for Tesla AI processing. Each of the two systolic arrays on a single FSD Chip are capable of 36 trillion operations per second, and there are two FSD Chips for redundancy.
HW3 cars are equipped with eight cameras, in the same locations and covering the same directions as the HW2 and HW2.5 cameras. Each of the eight cameras supplied with HW3 use the same AR0136A image sensor supplied by Onsemi, which has a maximum resolution of 1280×960 (1.2-megapixel) and a 3.75 μm pixel size. Initial versions of HW3 also included a Continental ARS4-B radar module.
https://en.wikipedia.org/wiki/Tesla_Autopilot_hardware
Full Self-Driving Chip (FSD Chip, previously Autopilot Hardware 3.0) is an autonomous driving chip designed by Tesla and introduced in early 2019 for their own cars. Tesla claims the chip is aimed at autonomous levels 4 and 5. Fabricated on Samsung’s 14 nm process technology, the FSD Chip incorporates 3 quad-core Cortex-A72 clusters for a total of 12 CPUs operating at 2.2 GHz, a Mali G71 MP12 GPU operating 1 GHz, 2 neural processing units operating at 2 GHz, and various other hardware accelerators. The FSD supports up to 128-bit LPDDR4-4266 memory.
The security system is designed to ensure the chip only executes code that has been cryptographically-signed by Tesla.
https://en.wikichip.org/wiki/tesla_%28car_company%29/fsd_chip
Not to mention, these chips are heavily co-designed with the @Tesla_AI software teams to achieve incredible performance, while also beating by a large margin on perf per Watt and perf per dollar.
For instance, AI4 can process and understand a million pixels of streaming video within ~1 ms. This is only achievable because the software AND the hardware are designed together to achieve this performance point.
https://x.com/aelluswamy/status/1992649962761670935?t=aRCG1DJjxMisOYPNrZLtYw&s=19