Athos Silicon: SoC with redundant UCIe chiplets for autonomous driving and more

The Mercedes-Benz spin-off Athos Silicon is working on functionally safe SoCs that will not only be used in autonomous vehicles with the three-pointed star.

listen Print view
White Mercedes-Benz stands on a path against a blue background

Unsurprisingly, Athos is testing its first Polaris in Mercedes-Benz vehicles. However, because the chiplets are still in the tape-out stage, only one emulator is currently running in the first of four vehicles.

(Image: Athos Silicon)

8 min. read
Contents

In the automotive world, technology for autonomous driving continues to open the checkbooks of car manufacturers, so established semiconductor companies such as Intel and Nvidia are keen to skim off the cash with automotive offshoots of their developments. However, their products, approaches and business practices do not always match the car manufacturers' expectations.

Mercedes-Benz has therefore been working on its own system-on-chip (SoC) at a research center in Silicon Valley in recent years. This is intended to meet both the performance requirements for level 3 and 4 autonomous driving as well as the regulatory requirements regarding reliability that driver assistance systems must satisfy.

However, the development of high-performance chips is expensive: with a forecast annual in-house requirement in the low single-digit million range, Mercedes alone cannot exploit the economies of scale that semiconductor developers are otherwise accustomed to in order to recoup the massive development costs of chips on modern production processes.

Six months ago, Mercedes therefore decided to spin off all of its previous efforts to the spin-off Athos Silicon – with a stake of only around 20 percent. The win-win situation: Athos already has an initial customer who is familiar with the product, but also has all the freedom to offer the technology to other interested parties and thus achieve higher unit numbers.

At the Hot Chips conference currently taking place in Stanford, California, Athos CTO Francois Piednoel explained the concept to us in more detail. Piednoel was a test and performance engineer at Intel for many years until he set off for new shores (Mercedes-Benz) seven years ago. His last project at Intel at the time was Skylake X aka Core i-7000X(E).

Videos by heise

Piednoel is also a hobby pilot and is now transferring a principle of functional safety from aviation to technology for cars. Athos uses several identical chiplets that work redundantly so that in the event of deviations, the majority principle decides which calculation result is used. Athos calls this approach Multiple Systems-on-Chip (mSoC).

The idea is initially – not nearly as strict as in avionics, partly due to less stringent regulatory requirements –. In passenger aircraft, three functionally identical computers with different architectures (x86, ARM, MIPS) are common, which work in parallel throughout and compare their results with each other. Athos relies on two identical chiplets (each with its own working memory) for the first processor, codenamed Polaris, and a third, smaller chiplet in the middle, which performs the comparison. The latter uses frame parameters and monitoring functions to decide which of the two large chiplets it believes if the two deliver different results.

Athos developed the chiplets in cooperation with DreamBig Semi (a group of former Marvell engineers). Communication takes place via the chiplet standard UCIe. Athos has supplemented this, for example with a side channel and the option of attaching different chiplets to the same host. The latter is based on the PCIe standard, where a long x16 slot (e.g. for graphics cards) can also accommodate and use a shorter x4 card. According to Piednoel, UCIe itself currently only provides for inflexible 1:1 connections.

Render image of a Polaris processor. Athos' chiplets are located in the top and bottom center, the DreamBig chiplet for decision making in the middle and DRAM components on the sides.

(Image: Athos Silicon)

Polaris should offer enough performance for Level 3 autonomous driving. The first chiplet has recently completed the so-called tape-out and the second is about to do so. This means that Samsung's manufacturing division has received the design and is now creating suitable exposure masks for production. Engineering samples for interested customers should be available at the beginning of 2026. The second stage, codenamed Northstar, which will then be suitable for level 4, is set to ignite just one year later. The highlight: Northstar uses the same chiplets as Polaris, but more of them, namely eight large ones and several small ones in between. As with Polaris, the partitioning remains a direct coupling of two chiplets that receive the same workload. Nevertheless, Northstar should deliver more raw computing power than Nvidia's Thor.

Athos takes over complete wafers from Samsung as a foundry and takes care of testing, sorting the chiplets according to production quality (binning) and other processing stages itself. One of the reasons for this is that Athos uses Silicon Box for packaging. The Singapore-based company offers a special technology: Silicon Box blows plasma between the die and the carrier in a vacuum and then presses the two components against each other at a defined speed. This causes the contact surfaces to weld together inseparably. As a result, the connections are mechanically more stable than soldered connections, where the solder balls can break off over time due to continuous vibrations (which inevitably occur when driving).

Athos wants to sell its processors not only directly to car manufacturers, but also to their suppliers. The AI models that run on them can in turn be customized. This means that, despite identical hardware, there is an individualization option that manufacturers can use to stand out from one another or trigger competition between suppliers.

Athos does not attach any importance to being publicly named as the manufacturer of the processors. Mercedes can sell the entire finished system as an in-house development under its own name and any other car manufacturer can do the same. This is a clear side-swipe at Nvidia's public image with regard to its drive platforms and is much more in line with industry practice. Usually, only experts know what exactly in a specific vehicle comes from the brand itself and what comes from Bosch or Continental or Valeo or ZF or whoever.

According to Piednoel, Athos is already in advanced talks with an Asian car manufacturer that is interested in Polaris and Northstar. He did not want to give a name when asked, but emphasized that it is not a Chinese manufacturer: he is not allowed to sell his technology to such manufacturers due to export restrictions.

Another scaling option that Athos is considering is expansion into other markets besides automotive. The technology used tracks small cubes in space, says Piednoel. For autonomous driving, the spatial resolution is around one hundred meters to the side and several hundred meters in front of and behind the vehicle, with a cube edge length of around two and a half centimeters.

Depending on the application, the edge length is an adjustment screw that Athos can turn without the need for other computing hardware. With smaller cubes, the technology is intended to be suitable for robots that do not have to navigate at highway speed but need to be particularly precise in changing environments.

If Athos enlarges the cubes, the company will end up with aircraft. Piednoel stated that he recently had to have his safety rating renewed because he was in talks with a drone manufacturer. The energy hunger of his computing platform (which he did not want to name specifically) does not play a role here: the specific project in question is not a handy aircraft with electric rotors, but has a turbine on board ... (mue)

Don't miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.