[ad_1]
Tesla will spend $1 billion on its Dojo supercomputer by means of the following yr because it beefs up video recognition capabilities for its autonomous automobiles.
“I believe we shall be spending one thing north of $1 billion over the following yr on – by means of the following yr, it’s properly over $1 billion in Dojo,” mentioned Elon Musk, Tesla’s CEO, throughout an earnings name.
The corporate additionally plans to deploy 300,000 Nvidia A100 GPUs by the tip of 2024, which is able to complement the Dojo supercomputer, which is able to deploy Tesla’s homegrown D1 chip.
“We could attain in-house neural web coaching functionality of 100 exaflops by the tip of subsequent yr,” Musk added.
Tesla makes use of a number of sorts of {hardware} for AI inferencing and coaching. It has on-car {hardware} working software program known as Autopilot, which is for assisted driving, and a beta model of Full Self-Driving (FSD), which is for hands-free autonomous driving.
Tesla mines its fleet for visible knowledge gathered by eight cameras and sensors, which it then feeds into its coaching mannequin. The present knowledge middle the place fashions are educated contains greater than 14,000 GPUs.
“Thus far, over 300 million miles have been pushed utilizing FSD beta. That 300-million-mile quantity goes to appear small in a short time. It’ll quickly be billions of miles, then tens of billions of miles,” Musk mentioned.
Tesla additionally began manufacturing of its Dojo supercomputer, which is a separate AI coaching cluster that has racks of homegrown D1 chips. The D1 delivers 22.6 teraflops of FP32 efficiency, has 50 billion transistors, 10TBps of on-chip bandwidth, and 4TBps of off-chip bandwidth.
Elon Musk. (COMEO/Shutterstock)
“That’s what Dojo is designed to do – optimize for video coaching. It’s not optimized for LLMs. With video coaching, you’ve gotten a a lot greater ratio of compute-to-memory bandwidth, whereas LLMs are typically reminiscence bandwidth,” Musk mentioned.
Tesla is rapidly gathering knowledge from FSD software program in its automobiles to create a system that can make autonomous driving secure and scale back on-road deaths.
“With a purpose to construct autonomy, we additionally want to coach our neural web with knowledge from tens of millions of autos … the extra coaching knowledge you’ve gotten, the higher the outcomes,” Musk mentioned, including “We see a transparent path to full self-driving being 10 instances safer than the common human driver”
However a scarcity of {hardware} is slowing down the video coaching. Musk particularly praised Nvidia CEO Jensen Huang, who was working to get Tesla extra GPUs to assist meet their computing want for computing pace.
“Frankly, I don’t know if they might ship us sufficient GPUs. We would not want Dojo, however they’ll’t. In order that they’ve received so many shoppers. They’ve been form sufficient to nonetheless prioritize a few of our GPU orders,” Musk mentioned.
Musk is betting on heavy R&D and AI expenditure to offer Tesla a aggressive benefit over rival automotive makers, which aren’t placing as a lot effort into coaching autonomous programs from scratch. Tesla this week reported second-quarter income of $21.27 billion, rising by 46% in comparison with the identical quarter the earlier yr.
The earnings have been thought of a disappointment because the working earnings was 9.6%, a year-over-year decline from 14.6%. That was partially attributable to decrease common promoting costs of Tesla’s electrical autos. Tesla supplied deep reductions on the S, X and three fashions in the latest working quarter.
Associated
[ad_2]
Source link
