TSMC may make the next self-driving Tesla chip

Our vehicles are getting smarter, and this is not such a superficial and forced statement as the one that manufacturers attach to every gadget without thinking: with a data connection, a lot of sensors and diverse functions, a large touch screen, an application store and games, as well as customizable profiles, driving assistance services and self-driving in the future, the Teslas for example, they have a very serious software background in addition to performance, comfort and security, with wireless updates coming for years, just like mobile phones.


Elon Musk has been promising for years, but his full self-driving is still not ready, and who knows when the driver can sit back in his Tesla, if at all. It is also not clear whether their camera-based solution, Waymo’s LiDAR and 3D map solution, or the solution relying on many more other vehicles and external sensors, will be successful. In addition to emergency braking and other safety solutions, depending on the self-driving package, Teslas are in any case able to drive onto the highway, maintain a lane, change lanes, overtake, cross onto another highway, and drive off. Even on low-grade roads and in cities, they roll to intersections and stop at lights and stop signs, as well as exiting and parking at a leisurely pace. However, full self-driving between two random points is still under development, and it is not certain that the computing capacity of the currently used chip, the Tesla HW3.0, will be enough to solve the problem on its own. The FSD beta program, on the other hand, already has more than a hundred thousand users in North America, with full urban (and rural) self-driving, but with the usual basic expectation of increased driver attention and immediate intervention. In addition to the fact that the fast-developing but still often buggy software has a lot to improve, the legal framework is also missing so that the driver can start watching a series on the screen, and then once the steering wheel and pedals are removed from the vehicles.

The current Tesla autopilot hardware, HW3.0. Source: Tesla


The FSD Beta is more widely available at version, and the internal team is already road-testing the main version 11, which combines the previously separate highway and city self-driving code. The neural network-based, self-learning software is refined by Tesla’s Nvidia GPU cluster-based computer park based on real camera and sensor data, as well as photorealistic simulation of rare and dangerous traffic situations, but the in-house developed Dojo supercomputer is already being prepared, the base unit of which, the Tesla D1 AI chip, is probably made by TSMC produced by system on wafer and fan out wafer based on technologies, on 7 sqm. As for the current vehicle’s self-driving computer (which is different from the on-board system with console-level graphics performance that AMD ships), the Tesla HW3.0 that debuted in 2019 is built on Samsung’s 14nm node, and the long-rumored HW4.0, which is based on a similar architecture but promises four times the computing capacity, can switch to TSMC’s 5 or 4 nm node, Taiwan’s UDN and China’s IT Home write independently.

Tesla's under-construction AI supercomputer, the ExaPOD based on D1 chips
Tesla’s under-construction AI supercomputer, ExaPOD, or Dojo, based on D1 chips (source: Tesla) [+]


The semiconductor giant may have received its largest order from a car manufacturer so far, which is not surprising given Tesla’s exponentially growing sales. By the way, TSMC’s automotive orders are constantly growing, and with the agreement, which has not been officially confirmed, Tesla can suddenly become the seventh largest customer. It is not yet known when the HW4.0 chips will end up in electric cars for the first time (there are those who believe that Tesla is already installing them, but the chances of this are quite small), and if the switch to full self-driving is necessary, the third vehicle in the existing fleet generation hardware to the fourth, will Tesla charge the cost of this to the owner. In any case, the brand installs all the cameras and hardware necessary for self-driving in all its manufactured vehicles, regardless of whether the customer pays for the extremely expensive FSD package.

This is costly from the production side, but there is a good reason: “full self-driving” (which is highly misleading and several organizations have spoken out against its use, let’s face it: rightfully so) is available to all owners immediately after purchasing the software package without a service visit. And the Tesla chips also operate in the background when the owner is operating his car – this is called shadow mode, and the Tesla team can use it to request driving data from a multi-million fleet to solve very rare cases (especially complex intersections, railway crossings, etc.) Elon Musk promises to open the FSD Beta in North America by the end of the year to anyone who bought the full package, with Europe to follow sometime later. But with the dates, Tesla is at war, to put it mildly, and the authorization of the FSD Beta may also depend on the region, especially in Europe. Whether it is a good idea to provide a one-time civilian with a software under development and then use it on public roads is a good basis for discussion. Although beta requires constant attention, failing which the vehicle will pull over and exclude the driver from the program after three warnings, the owner is legally responsible for any accident caused by the FSD that the driver did not prevent. For the time being, there have been no FSD Beta accidents involving hundreds of thousands of users that caused personal injury, but now the accident-prone winter is coming, with longer nights and more slippery roads. It will be interesting to compare the accident statistics of the system with the average in the long term.

Tesla Semi prototype at Megacharger charging station
Tesla Semi prototype at Megacharger charging station (source: Electrek) [+]

In any case, there is no doubt that the software background of a vehicle – and especially an electric car – is at least as important as the attention paid to comfort, safety or service life, and the robotaxi market has tens of billions of forints of potential. So it really doesn’t matter who, when, for how much and with how many vehicles can introduce this on the market, which can be affected by regulatory obstacles in addition to technological challenges. Hundreds of small startups and large companies worldwide are interested in this, and no one has a complete recipe yet. On the other hand, Tesla’s electric tractor unit is ready to go, with a real-life driver, of course: the EV manufacturer will hand over its first trucks to Pepsi in a week, on December 1st, and in addition to the vehicle, the Megacharger’s high-performance fast charger can be demonstrated at the invitation-only event.

It would go in the Winged Bounty Hunter, but how will the Cybertruck paint on public roads
It would go in the Winged Bounty Hunter, but how will the Cybertruck paint on public roads (source: Tesla) [+]

Speaking of which, self-driving and the domestication of trucks seem practical, after all, the lack of drivers alone causes supply disruptions worldwide. For now, however, it does not seem that HW4.0 will debut in the tractor unit, or that some kind of self-driving will be available from the start. It is more likely that the chip produced at TSMC will start next year with the Cybertruck electric van, which has not become less divisive since its announcement in 2019 with its brutalist stainless steel exoskeleton exterior. In this respect, Tesla is typical: every page has a definite opinion about it, and Elon Musk on Twitter is also trying to increase the number of critics exponentially.

Back to top button