Tesla says AI5 chip leap

Tesla’s next self-driving chip has reached the “tape-out” stage, meaning Tesla has finished the design and sent AI5 to manufacturing partners. (electrek.co) Elon Musk said on April 15 that Tesla’s chip team had taped out AI5 and thanked Taiwan Semiconductor Manufacturing Company and Samsung for helping bring it to production. Tesla North reported Musk also said AI6 and Dojo3 are already in development. (teslanorth.com) A tape-out is the point where a chip design stops being a computer file and starts becoming physical silicon in a factory. It is a milestone, not mass production, and Electrek reported vehicle-scale output is still more than a year away. (electrek.co) Tesla has been building its own driving computers for years instead of buying all of that silicon from Nvidia. The company told investors in April 2019 that its Full Self-Driving computer was already in production, making custom chips part of Tesla’s autonomy plan well before AI5. (ir.tesla.com) The current generation in Tesla vehicles is commonly called AI4 or Hardware 4, and AI5 is the successor Musk is now pitching for cars, robots and Tesla’s computing clusters. Tesla North said Musk has previously described AI5 as up to 10 times more capable than the hardware now used in Model 3 and Model Y. (teslanorth.com) The comparison driving the latest debate comes from Musk’s own posts, relayed by TechPowerUp, that put two AI5 chips against one Nvidia Blackwell processor and a single AI5 around Nvidia Hopper-class performance. Those are company claims, not independent benchmark results. (techpowerup.com) Nvidia’s Blackwell is a data-center processor built for training and running large artificial intelligence models in server racks, not inside a car. Nvidia says the Blackwell platform is designed for trillion-parameter large language models and real-time generative artificial intelligence on a data-center scale. (nvidia.com) Tesla’s version of the problem is different: it wants a chip that can run neural networks inside a vehicle, on automotive power and thermal limits, while also scaling to robots and server clusters. Tesla North reported analysts expect the custom silicon to fit in a vehicle computer while using far less power than a rack-scale graphics processor. (teslanorth.com) Manufacturing is also part of the story. TechPowerUp reported Tesla plans to split AI5 production between Samsung’s plant in Taylor, Texas, and Taiwan Semiconductor Manufacturing Company’s Arizona facility, instead of relying on a single foundry. (techpowerup.com) What happens next is less dramatic than the numbers in Musk’s posts: first silicon has to come back from the factory, then Tesla has to validate the chip, fix bugs if needed and ramp production. The news this week is that AI5 is now past the design stage and into the long manufacturing pipeline. (electrek.co)