Tesla has spun up the world’s fifth most powerful supercomputer, which will be used to train the neural nets powering the Autopilot and the AI that will be used for its Full Self-Driving software. The supercomputer is actually less powerful than its planned Dojo supercomputer, which will be capable of a full exaflop, or 1018 operations per second.
Elon Musk has previously said that Dojo is likely to be ready by the end of this year. Until then, Tesla’s Autopilot and Full Self-Driving programs rely on its existing computing power. This most recent supercomputer is the third one that it has created in-house and was unveiled by the company’s AI chief, Andrej Karpathy, during a presentation at the 2021 Conference on Computer Vision and Pattern Recognition. According to the presentation, this supercomputer’s specs include:
- 720 nodes of 8x A100 80GB. (5760 GPUs total)
- 1.8 EFLOPS (720 nodes * 312 TFLOPS-FP16-A100 * 8 gpu/nodes)
- 10 PB of “hot tier” NVME storage @ 1.6 TBps
- 640 Tbps of total switching capacity
“For us, computer vision is the bread and butter of what we do and what enables Autopilot. And for that to work really well, we need to master the data from the fleet, and train massive neural nets and experiment a lot,” Karpathy said during the presentation.
Tesla vehicles have sent back enough real-time driving data to make a solid start on that, with more than a billion miles driven since its electric vehicles started hitting the road. The Autopilot and Full Self-Driving engineering team were working on labeling of objects seen through the onboard vehicle camera systems. That labeling is currently being done by humans. The software can already recognize most road signs, turn lanes, and lane markings on the road.
Despite some privacy concerns, the transmission of data to Tesla’s supercomputers helped investigators determine that the Autopilot was not engaged at the time of a recent fatal wreck in Texas despite authorities initially speculating that it was. Data from Tesla’s cameras have also cracked criminal cases like a string of shootings with a BB gun in California and a series of hate crimes against members of a primarily African-American church in Missouri.
Besides using Dojo’s predecessors to train Tesla’s neural nets, Elon Musk has indicated that it would be willing to lease their processing power to other companies to train their own neural nets and AI programs. Karpathy didn’t elaborate much on progress made on Dojo or what its final specs would wind up being besides saying that the upcoming supercomputer would be optimized for neural nets.
According to communications between Tesla employees and the DMV, development of Tesla’s Autopilot and Full Self-Driving software is likely to take longer than Elon Musk seems to expect. Tesla’s website includes the disclaimer that drivers should remain alert at all times while using the software and has recently backed that up by activating a camera mounted to the rear-view mirror. The latter may have been sparked by recent incidents in which California residents were seen pulling stunts like riding in the back seat of a driverless Tesla vehicle with Autopilot engaged.
These incidents were blamed on people being dumb rather than any allegedly overstated capabilities of Tesla’s software. Tesla is continuing to improve on the software with the use of supercomputers to train the neural nets and AI that power them and Full Self-Driving itself is still in beta.