Paul van Gerven
25 April 2019

Tesla designed “the best chip in the world” for autonomous driving at its first try. Really?

We have known since 2017 that Tesla was working on its own self-driving hardware. Last October, Elon Musk said it would be available in six months’ time. And, in an unusual example of the high tech visionary correctly predicting the timing of a Tesla product launch, the hardware has indeed arrived. At the Tesla Autonomy Investor Day earlier this week, the company announced it has been shipping cars with in-house developed AI chips that will eventually enable (almost) fully autonomous driving.

At the event, Musk called the IC “the best chip in the world … by a huge margin” for self-driving capability. Considering it was the company’s first attempt at IC design, that would be a remarkable achievement. Let’s have a closer look and see what’s what.

What kind of chip are we talking about?

The 260 mm2 14nm chip, fabbed by Samsung, contains 6 billion transistors divided into CPU, GPU, memory, video and two neural network accelerator blocks. Only the latter blocks are home-grown; the other parts are licensed. Together, the neural accelerators, clocked at 2 GHz, can process up to 1 TB of data per second and perform 72 tera operations per second (TOPS). This results in enough processing power to analyze 2,100 video frames per second, which are captured by eight cameras onboard a Tesla (in addition to other sensor data, but not from lidar – Musk is not a fan).


Device lifecycle management for fleets of IoT devices

Microchip gives insight on device management, what exactly is it, how to implement it and how to roll over the device management during the roll out phase when the products are in the field. Read more. .

Who designed it?

Semiconductor industry veterans Pete Bannon and Jim Keller. On paper, they certainly have the credentials to design something revolutionary. Bannon worked at Intel and then at PA Semi, which was later bought by Apple when it decided to design its own Arm-based processor lines. The chip architect joined Tesla in 2016. Keller worked on Apple’s A4 and A5 processors before moving to AMD, where he directed the development of the Zen architecture. Keller also joined Tesla in 2016.

Tesla board_web
Photo: Tesla

How does Tesla’s chip stack up against the competition?

During the event, Tesla compared its chips with NVidia hardware, which Tesla had been using before in its Autopilot functionality. Since Tesla’s system uses two chips side by side, it performs 144 TOPS, while NVidia’s Xavier does 21 TOPS, said Tesla. The power usage was said to compare favorably as well.

There’s a lot wrong with this comparison. First of all, the Xavier is capable of 30 TOPS. Secondly, the Xavier is aimed at assisted, not fully autonomous driving. The fair comparison would be with NVidia’s Drive AGX Pegasus, which is reported to perform 320 TOPS. And thirdly, Tesla uses two chips for redundancy. They perform the same calculations and then compare notes, which means you cannot just add up the TOPS.

In fact, any comparison is troublesome. Tesla’s solution is designed specifically for the sensor hardware that the company has already committed to, while NVidia’s chips have to work with whatever configurations their car maker clients prefer. So, indeed, Tesla’s best chip in the world is the best … for Tesla.

Still, NVidia did offer praise for Tesla, saying it had “raised the bar for self-driving computers.”

Great, now where’s my robotaxi?

The media have been reporting that Musk promised a million fully autonomously driving Teslas would hit the road by the end of next year. That isn’t quite what he said: he said a million Teslas capable of autonomous driving would be on the road, ie with the required hardware installed. Step by step, firmware update after update, the cars would then move towards self-driving capability. Tesla has by no means proven that the presented hardware is, in fact, capable of powering autonomous driving, though. Nor would anyone call the development of the required software a triviality. So it might take a while before you hail your first robotaxi.