Maria Montero

Tesla Announces Creation of ‘World’s Best Chip’ for …

On his “Autonomy Day” today, Tesla He detailed the new custom chip that will run the self-driving software in their vehicles. Elon Musk, quite peremptorily, called it “the best chip in the world … objectively.” That might be a stretch, but it should definitely get the job done.

Called for now the “fully autonomous computer,” or the FSD computer, it is a high-performance, special-purpose chip built (by Samsung, Texas) exclusively with autonomy and safety in mind. If and how it really outperforms its competitors is not a simple question and we will have to wait for more data and more detailed analysis to say more.

Former Apple chip engineer Pete Bannon reviewed the FSDC specifications, and while the numbers may be important to software engineers working with the platform, what is more important at a higher level is meeting various specific requirements for self-driving tasks.

Perhaps the most obvious feature for AVs is redundancy. The FSDC consists of two mirrored systems side by side on one board. This is an important, albeit almost unprecedented option, simply because dividing the system in two naturally also divides its power, so if performance were the only metric (if it were a server, for example), it would never I would.

Here, however, redundancy means that if a bug or corruption creeps in some way or another, one of the two systems will be isolated and the reconciliation software will detect and flag it. Meanwhile, the other chip, in its own power and storage systems, should not be affected. And if something happens that breaks both at the same time, the architecture of the system is the least of your worries.

Redundancy is a natural choice for AV systems, but is made more acceptable by the extreme levels of acceleration and specialization that are possible today for neural network-based computing. A normal general-purpose CPU, like the one you have in your laptop, will be trained by a GPU when it comes to graphics-related calculations, and similarly, a special neural network computing unit will outperform even a GPU. As Bannon points out, the vast majority of calculations are a specific mathematical operation and one that brings huge performance benefits.

Combine that with high-speed RAM and storage and you’ve got very little bottleneck when it comes to running the more complex parts of automated systems. The resulting performance is impressive, enough to make a proud hint of musk during 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 happened. Not better by a small margin, better by a large margin.”

Let’s take this with a little salt, as surely the engineers at Nvidia, Mobileye, and other self-driving concerns might disagree with the statement for some reason or another. And even if it is the best chip in the world, there will be a better one in a few months, and regardless of that, the hardware is only as good as the software that runs on it. (Fortunately, Tesla has incredible talent on that side, too.)

(A quick note for a piece of terminology you may not be familiar with: OPs. This is short for operations per second, and is measured in billions and billions these days. FLOPs is another common term, which means operations Second, these refer to the higher precision mathematics that supercomputers often use for scientific calculations.One is not better or worse than the other, and they should not be directly compared or considered interchangeable.

High-performance computing tasks tend to drain your battery, like doing transcoding or HD video editing on your laptop, and sting the dust after 45 minutes. If your car did that, you would be angry, and rightly so. Fortunately, a side effect of acceleration tends to be efficiency.

The full FSDC runs on about 100 watts (or 50 per compute unit), which is pretty low; It’s not a low-end cell phone chip, but it’s well below what a high-performance desktop or laptop would get – less even than many individual GPUs. Some AV-oriented chips attract more, others draw less, but Tesla’s claim is that they get more power per watt than the competition. Again, these claims are difficult to examine immediately considering the closed nature of AV hardware development, but it is clear that Tesla is at least competitive and may very well beat its competitors on some important metrics.

Two more specific AV functions found on the chip, although not duplicated (the computational paths converge at some point), are some CPU lock jobs and a layer of security. Lockstep means that it is being very carefully applied that the time on these chips is the same, ensuring that they are processing the exact same data at the same time. It would be disastrous if they got out of sync with each other or with other systems. Everything in AV relies on very precise timing and minimizes lag, so robust blocking measures are in place to keep it in order.