Tesla Is Building Its Own AI Chips For Self-Driving Cars (techcrunch.com)

Yesterday, during his quarterly earnings call, Tesla CEO Elon Musk revealed a new piece of hardware that the company is working on to perform all the calculations required to advance the self-driving capabilities of its vehicles. The specialized chip, known as "Hardware 3," will be "swapped into the Model S, X, and 3," reports TechCrunch. From the report: Tesla has thus far relied on Nvidia's Drive platform. So why switch now? By building things in-house, Tesla say it's able to focus on its own needs for the sake of efficiency. "We had the benefit [...] of knowing what our neural networks look like, and what they'll look like in the future," said Pete Bannon, director of the Hardware 3 project. Bannon also noted that the hardware upgrade should start rolling out next year. "The key," adds Elon "is to be able to run the neural network at a fundamental, bare metal level. You have to do these calculations in the circuit itself, not in some sort of emulation mode, which is how a GPU or CPU would operate. You want to do a massive amount of [calculations] with the memory right there." The final outcome, according to Elon, is pretty dramatic: He says that whereas Tesla's computer vision software running on Nvidia's hardware was handling about 200 frames per second, its specialized chip is able to crunch out 2,000 frames per second "with full redundancy and failover." Plus, as AI analyst James Wang points out, it gives Tesla more control over its own future.

Comments

Different headline than I expected

[DontBeAMoran]

From the previous thread about Tesla, I expected this headline to read "Tesla is now building their own arcade cabinets".

Re: Different headline than I expected

[ShanghaiBill]

Tesla has invented the fpga?

This is not an FPGA. It is a matrix math engine, like Google's TPU.

They do not own a fab

Other than Intel, nobody owns their own fabs anymore.

You just code up your chip in Verilog, debug in a simulator, and upload it to TSMC.

or have the cpu designers that Intel, amd and invidea have.

Neither did Google, but their TPU is a big success.

Maximum Overdrive 2: Revenge of the Dissed

[mykepredko]

You know if you've ever said anything nasty about Elon.

Now, his vehicles know.

Be afraid. Be very, very afraid.

Building proprietary silicon could be dangerous

[postbigbang]

For better and worse, keeping things proprietary means it's by definition both closed source, and tested only to one's own environment. Although it produces fast yields, it doesn't have many eyes. Many eyes and many hours are needed to vet the integrity and edge cases (like cliff edges) before safety can be assured.

It's a risky, expensive, and proprietary endeavor. If everyone (systems builders) were using similar development, the testing age could be completed in a concurrent time, rather than a serial/iterative time. I'm betting against this turning out well.

Re:Building proprietary silicon could be dangerous

[religionofpeas]

Neural net calculations are pretty simple, just repeated many times over. Testing the silicon should be relatively simple compared to general purpose CPU or even GPU design.

Re:Building proprietary silicon could be dangerous

[RhettLivingston]

Every NN is proprietary, and that is where the functionality to worry about is at. The performance on "edge cases" in driving is directly related to how much compute power you can throw at it. Tesla is multiplying its compute power. The edge cases will improve. Staying with the general purpose GPU instead of true NN hardware will guarantee continued unhandled edge cases.

This HW is undoubtedly also more energy efficient. That is the real key. They could stack on more boards, but these units are already consuming a significant amount of the vehicle's energy. The trick is to get more compute power with the same or less energy. NN specific HW is going to be a requirement to have that happen.

Everyone in the industry has known that GPUs will not be used past the first generation or so. They are development HW. Someone will eventually come up with a general purpose NPU that will win the market, but it hasn't happened yet - mostly because NN implementations haven't settled.

If anyone needs me...

[dfn5]

I'll be in the basement hiding from the cars

Surely a bad decision

[mugurel]

"We had the benefit [...] of knowing what our neural networks look like, and what they'll look like in the future,"

Really? If they take their neural network development seriously I don't think they know what their networks will look like in ten years. It's a research area in the middle of a transformation. Using architectures molded into hardware is probably just costly and will act as an antagonist to innovation. I don't think having 2000 vs 200 frames per second right now outweighs that downside.

To what end?

[Kjella]

A car travelling at 90 mph is moving about 4 cm/millisecond. So going from 200 fps to 2000 fps is going from 20 cm to 2 cm per cycle. What's the use of recognizing a car every two centimeters? For a jogger at 9 mph it's down from 2 cm to 2 mm. It's neat and all but I don't see how that necessary to react in the time frames a car needs to react. Even if it takes 3-4 frames for the car to get a motion vector 0.2 seconds is still way quicker than a human and 0.02 seconds doesn't bring that much.

Re:To what end?

[TomGreenhaw]

I bet that this allows them to have more cameras. 2000 fps for one camera could be 250fps for 8 cameras. it could also be used for much higher resolution cameras that have fisheye or insect like lenses.

Re:To what end?

[dgatwood]

Maybe after a point, but up until that point, the main risk is reacting too slowly. Ask anybody with an AP2 Tesla how well it handled curves prior to earlier this year. Of they don't use the word "lag", they don't know software, and if their eyes don't bug out in abject terror, they don't know how to drive.

Basically, it had (and still has, to a lesser extent) trouble with lane keeping, because its reactions lagged behind reality, and it started turning way too late, resulting in uncomfortable turns, getting dangerously close to barriers and center lines, etc. This is better in current versions, but I still get scared enough to take manual control a couple of times per day.

So right now, performance is still their main problem. This is a very welcome announcement.

Re:To what end?

[philmarcracken]

That's the answer they want people to hear. The real answer is no longer having to pay nvidia.