Artificial IntelligenceAutomotiveRoboticsTransportation

Tesla vaunts creation of ‘the best chip in the world’ for self-driving

At its “Autonomy Day” today, Tesla detailed the brand-new custom chip that will be running the self-driving app in its vehicles. Elon Musk rather peremptorily named
it “the best chip in the world…objectively.” That might be a stretch, but it certainly should get the job done.

named
for now the “full self-driving computer,” or FSD Computer, it is a high-performance, unique-purpose chip built (by Samsung, in Texas) solely with autonomy and safety in mind. Whether and how it actually outperforms its competitors is not an uncomplicated request and we will have to wait for more data and closer analysis to say more.

Former Apple chip engineer Pete Bannon went over the FSDC’s specs, and while the numbers may be important to app engineers working with the platform, what’s more important at a higher stage is meeting various requirements accurate to self-driving tasks.

Perhaps the most obvious feature catering to AVs is redundancy. The FSDC consists of two duplicate systems right next to each other on one board. This is a significant decision, though hardly unprecedented, simply because splitting the system in two naturally divides its energy as well, so if performance were the only metric (if this was a server, for example) you’d never do it.

Here, however, redundancy means that should an error or destruction creep in somehow or another, it will be isolated to one of the two systems and reconciliation app will detect and flag it. Meanwhile the other chip, on its own energy and storage systems, should be unaffected. And if something happens that breaks both at the same moment, the system architecture is the least of your worries.

Redundancy is a natural decision for AV systems, but it’s made more palatable by the extreme stages of acceleration and specialization that are feasible nowadays for neural network-based computing. an orderly general-purpose CPU like you have in your laptop will get schooled by a gpu when it comes to graphics-related calculations, and similarly a unique compute unit for neural networks will beat even a gpu. As Bannon notes, the vast majority of calculations are an accurate math operation and catering to that yields huge performance benefits.

Pair that with high speed RAM and storage and you have very tiny in the route of bottlenecks as far as running the most complex parts of the self-driving systems. The resulting performance is great, enough to make a proud Musk chime in during the presentation:

“How could it be that Tesla, who has never designed a chip before, would design the best chip in the world? But that is objectively what has occurred. Not best by a tiny margin, best by a gigantic margin.”

Let’s take this with a grain of salt, as surely engineers from Nvidia, Mobileye, and other self-driving concerns would take issue with the statement on some grounds or another. And even if it is the best chip in the world, there will be an acceptable one in a few months — and regardless, hardware is only as good as the app that runs on it. (Fortunately Tesla has some unbelievable talent on that side as well.)

(One quick note for a piece of terminology you might not be familiar with: OPs. This is short for operations for second, and it’s measured in the billions and trillions (TOPs) these days. FLOPs is another common term, which means floating-point operations per second; these pertain to higher-precision math often used by supercomputers for scientific calculations. One isn’t acceptable or worse than the other, and they shouldn’t be compared directly or considered exchangeable.)

Update: Right on cue, Nvidia objected to Tesla’s comparison in a statement, calling it “inaccurate.” The Xavier chip Tesla compared its hardware favorably to is a more lightweight chip for autopilot-type features, not full self driving. The 320-TOP steer AGX Pegasus would have been an acceptable comparison, the company said — though admittedly the Pegasus pulls about four times as much energy. So per-watt Tesla comes out ahead by the stats we’ve seen. (Chris here named
it during the webcast.
)

High-performance computing tasks tend to drain the battery, like doing transcoding or HD video editing on your laptop and it bites the dirt after 45 minutes. If your vehicle did that you’d be crazy, and rightly so. Fortunately a side effect of acceleration tends to be efficiency.

The whole FSDC runs on about 100 watts (or 50 per compute unit), which is beautiful low — it’s not cell phone chip low, but it’s well below what a desktop or high performance laptop would pull, less even than many solo GPUs. Some AV-oriented chips draw more, some draw less, but Tesla’s bay is that they’re getting more energy per watt than the tournament. Again, these claims are strenuous to vet immediately considering the closed nature of AV hardware development, but it’s clear that Tesla is at least competitive and may very well beat its competitors on some important metrics.

Two more AV-accurate features found on the chip, though not in duplicate (the compute pathways converge at some point), are some CPU lockstep work and a security layer. Lockstep means that it is being very carefully enforced that the timing on these chips is the same, ensuring that they are processing the accurate same data at the same moment. It would be devastating if they got out of sync either with each other or with other systems. Everything in AVs depends on very accurate timing while minimizing delay, so robust lockstep measures are put in place to keep that straight.

The security portion of the chip vets commands and data cryptographically to watch for, essentially, hacking attempts. Like all AV systems, this is a finely-oiled device and interference must not be allowed for any reason — lives are on the line. So the security piece watches the input and output data carefully to watch for anything suspicious like spoofed visual data (to ruse the vehicle into inference there’s a wayfarer, for example) to tweaked output data (say to prevent it from taking proper precautions if it does detect a wayfarer).

The most great part of all might be that this whole custom chip is backwards-compatible with existing Teslas, able to be dropped right in, and it won’t even cost that much. Exactly how much the system itself costs Tesla, and how much you’ll be charged as a customer — well, that will probably vary. But despite being the “best chip in the world,” this one is relatively affordable.

Part of that might be from going with a 14nm fabrication process rather than the sub-10nm process others have chosen (and to which Tesla may eventually have to migrate). For energy savings the smaller the acceptable and as we’ve established, efficiency is the name of the game here.

We’ll know more once there’s a bit more objective — truly objective, apologies to Musk — testing on this chip and its tournament. For now just know that Tesla isn’t slacking and the FSD Computer should be more than enough to keep your version 3 on the street.

Source
TechCrunch
Tags

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Close