Tesla unveils custom chip for AI-training supercomputer

August 29, 2021 // By Rich Pell
Tesla unveils custom chip for AI-training supercomputer
Electric vehicle company Tesla has designed a custom chip for training artificial intelligence (AI) networks in data centers.

The D1 chip, announced at the company's live-streamed AI Day event, is part of the company's previously announced Dojo neural network training supercomputer system, which is being developed to "process truly vast amounts of video data," such as video feeds from Tesla vehicles. Based on a 7-nanometer manufacturing process and optimized for machine learning workloads, the D1 chip features 362 teraflops of processing power.

The chips are said to comprise over 50 billion transistors and have a die size of 645 mm 2. The company says that 25 of these chips are placed onto a single "training tile," 120 of which are then combined across several server cabinets, resulting in over an exaflop of power.

Ganesh Venkataramanan, senior director of Autopilot hardware at Tesla told AI Day event particpants and viewers, "We are assembling our first cabinets pretty soon."

The Tesla technology, says the company, will be the fastest AI-training computer.

As for when the Dojo supercomputer will be ready, "We should have Dojo operational next year," says Tesla CEO Elon Musk.

Tesla

Related articles:
Nvidia unveils CPU for giant-scale AI and HPC
Microsoft announces new supercomputer, large-scale AI vision
Industry group calls for coordinated steps in autonomous driving development

Volkswagen plans own chip development

Mobileye gives a glimpse of its future technology

“We need standardized criteria for autonomous driving”


Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.