During the "Autonomy Day" today, Tesla details the new consumer chip that will use the self-management software in its vehicles. Elon Musk called it "the best chip in the world … objectively." This may be stretching, but it certainly has to do the job.
Currently called a "full computer for self-management" or FSD Computer, it is a highly productive, dedicated chip created (by Samsung, in Texas) solely for the sake of autonomy and safety. Whether and how it actually outperforms its competitors is not just a question and we will have to wait for more data and more in-depth analysis to say more.
Former Apple Chip Engineer Pitt Bannon has reviewed FSDC's specifications, and although the figures may be important for software engineers working with the platform, the more important thing at a higher level is to meet different requirements specific to self- management.
Perhaps the most obvious feature of audiovisual services is the abbreviation. FSDC consists of two duplicate systems side by side on one board. This is a significant choice, though it is hardly unprecedented, simply because splitting the system into two naturally splits its power, so if performance was the only indicator (if it was a server, for example), you will never do it.
Here, however, the surplus means that if any error or breakthrough breaks in any way, it will be isolated to one of the two systems, and reconciliation software will detect and mark it. At the same time, the other chip, on its own power and storage system, should be unaffected. And if something happens that both disrupts both, system architecture is the least worry.
Surplus is a natural choice for AV systems, but it's made more delicious than the extreme acceleration and specialization levels that are now possible for neural network computing. A regular general processor that you have in your laptop will be trained by a graphics processor when it comes to graphical calculations, and in the same way a special computing unit for neural networks will even beat the GPU. As Banang notes, most of the calculations are a specific mathematical operation and take care of what gives enormous benefits to performance.
Pair this with high speed RAM and storage and you have very little to overload as far as the most sophisticated parts of self-management systems are running. The resulting performance is impressive enough to make a proud musical cavalier during the presentation:
"How can it be that Tesla, who has never designed a chip, will design the best chip in the world? But it is objectively what happened. It is not best with a small margin, best with a large margin.
Let's take that with salt, as Nvidia, Mobileye engineers and other concerns that concern themselves, would question the statement for some reason. And even if this is the best chip in the world, it will be better for a couple of months – and whatever hardware is as good as the software that works with it. (Fortunately, Tesla has some amazing talents on this side as well.)
(A quick note for some terminology that you may not be familiar with: OP This is an acronym for operations per second and is measured in billions and trillions these days FLOPs is another general term that means floating point operations Second, they refer to more precise mathematics, which is often used by supercomputers for scientific computing. One is not better or worse than the other, and should not be directly compared or considered to be substitutable.
High-performance computing tasks tend to drain the battery by transcoding or editing HD video on your laptop, and after 45 minutes bites the dust. If your car did this, you would be crazy and right. Fortunately, the side effect of the acceleration tends to be efficient.
The entire FSDC is running at around 100 watts (or 50 per unit of computation), which is quite low – it's not a cell phone chip, but it's well below the desktop or high-performance laptop, which is less than many single GPUs. Some AV-oriented chips attract more, some decrease, but Tesla's claim is that they get more watts than the competition. Again, these claims are difficult to look at, given the closed nature of AV hardware development, but it is clear that Tesla is at least competitive and can beat its competitors on some important benchmarks.
Two more AV-specific chip functionalities, although not duplicated (computing paths converge at some point), are some processor operations and a protective layer. Lockstep means that it is applied very carefully that the time of these chips is the same, ensuring that they handle exactly the same data at the same time. It would be detrimental if they do not synchronize with each other or with other systems. Everything in the audiovisual devices depends on the very precise synchronization, while minimizing the delay, so that stable measures are introduced to preserve this.
The chip security section checks cryptographic commands and data to monitor, in essence, hacking attempts. Like all audio-visual systems, it's a finely-crushed machine and interference must not be allowed for any reason – life is on the line. Therefore, the security section carefully monitors the input and output data to monitor for suspicious behavior, such as false visual data (to make the car think that there is a pedestrian, for example) to correct the source data (say, to prevent taking appropriate action, finding a pedestrian).
The most impressive part of anything can be that the entire consumer chip is backward compatible with the existing Teslas that can be placed right here, and it will not cost even that much. How much does Tesla's body cost and how much you will be charged as a customer – well, it will probably vary. But although it is "the best chip in the world", it is relatively inexpensive.
Some of this may be due to the process of producing 14nm, not the sub-10nm process that others have chosen (and to which Tesla may eventually migrate). To save energy, the less, the better, and as we have found, efficiency is the name of the game here.
We will find out more when there is a more objective – really objective, excuses of Musk – testing this chip and its competition. So far, I just know Tesla is not slowing down and the FSD's computer must be more than enough to keep your model 3 on the road.