Tesla aims for new neural net computer in production in 6 months, results in 500-2000% increase in ops/sec, says Elon Musk
– Oct. 16th 2018 6:15 am ET
They are now aiming for the new computer to be in production in about 6 months and it could result in a 500-2000% increase in operation per second, according to Musk.
The release of this new computer with Tesla’s own AI chip would be the culmination of a long project that Tesla started about 3 years ago as it anticipated a need for more computing power in its vehicles.
Back in 2016, we first reported on Tesla quietly hiring legendary chip architect Jim Keller from AMD and we were fairly excited by the implications of Tesla hiring such an important chip architect.
At the time, we speculated that Tesla could be looking into making its own silicon at some point – speculation that was further reinforced after Keller’s hiring was followed by a team of chip architects and executives from AMD also joining Tesla.
Finally, our suspicions were confirmed two years later when Elon Musk confirmed that Tesla is working on its own AI chip.
Keller has since left Tesla, but the team is now led by Pete Bannon, who was amongst the other chip architects hired by Tesla over 2 years ago.
This summer, Bannon confirmed that the new chip will be released in a new ‘hardware 3’ suite for Autopilot which will take the form of a new computer that will replace the existing computer in the vehicles with Autopilot hardware 2.0 and 2.5 for those who ordered the Full Self-Driving Capability Package.
Timeline and Availability
At the time, Bannon confirmed that Tesla’s computer chip optimized for neural nets has been produced and they are aiming to release it next year.
Now Musk updated the timeline to “~6 months”
It means that they aim for the computer to enter production around the second quarter of 2019.
While it’s being introduced to new cars in production, Tesla will also need to produce a ton of computers for retrofits since there will likely be hundreds of thousands of cars eligible for the computer upgrade at that point.
Every owner who ordered the Full Self-Driving Capability Package, which has been available (to order) since October 2016, will receive the upgrade. The package has cost between $3,000 and $5,000 over the years.
In order to get the package, buyers need to have the ‘Enhanced Autopilot’ package, which costs $5,000 when ordering the vehicle and $6,000 if buying later through an over-the-air update.
Once you have this package, the ‘Full Self-Driving Capability’ costs $3,000 when ordering and it used to cost $4,000 when buying over-the-air later until Tesla increased the price to $5,000 in June.
Tesla told Electrek that they will honor the original price for existing owners who didn’t buy the package thinking that they could later upgrade over-the-air for $4,000, but it has yet to fix the issue for many buyers.
Musk previously said that Tesla’s hardware 3 will be able to handle 2,000 frames per second with redundancy.
He explained that they achieved that by building the new chip from the ground up to act as a ‘neural network accelerator’ based on the neural net that Tesla’s AI and vision team have been building.
The team deployed a new neural net with the recent version 9 software update.
Yesterday, we shared an analysis from a deep learning expert who took a look at the neural net and estimated to be significantly bigger than the previous version.
Musk confirmed that it features an about “400% increase in useful ops/sec due to enabling integrated GPU and better use of discrete GPU.”
Those GPUs are from Nvidia on the Autopilot 2.0+ hardware suites.
The CEO said that Tesla’s neural net computer upgrade should result in another 500% to 2,000% increase:
Tesla’s end goal is to power a computer vision neural net powerful enough to drive the car in any condition.
While they are making some significant progress, Musk noted that they still have “a long way to go.”
Featured Image: Look inside Tesla’s onboard Nvidia supercomputer for self-driving by Kyle Day