Athos Scraps Multi-Vendor Roadmap, Plans Chiplet Tape-Out

Six months after officially spinning out of Mercedes-Benz, functional safety chiplet startup Athos Silicon has redrawn its technology roadmap and will now rely on a purpose-built, in-house-designed chiplet, the company’s founders told EE Times in a recent interview.

The Athos team had been studying chip design for functional safety in autonomous driving for five years as part of Mercedes-Benz, and had spun out Athos with substantial IP at the end of 2024. The process was officially completed in October 2025. The plan had been a multi-chiplet SoC designed for functional safety with three compute SoC dies from a third-party supplier, DreamBig’s chiplet hub in the center, plus an NPU die from another third-party supplier.

DreamBig’s recent acquisition by Arm meant going back to the drawing board on hardware, Athos CTO François Piednoël told EE Times.

“The hub idea was so great, Arm acquired it for $240 million,” Piednoël joked.

Athos decided it couldn’t afford to wait for the acquisition dust to settle and that controlling its own chiplet design was, in any case, the way to go. The new design is based on a single SoC chiplet designed by Athos with third-party CPU, GPU, and NPU IP.

“This is how Silicon Valley is supposed to work—fail fast, adjust, and go again,” Piednoël said. “That plan was going to work, but now, since we have to pay for the chiplet, we can only pay for one. So we had to redo the design to be able to compete on pricing.”

Even given the extremely strict technical requirements, chip sales to automotive OEMs are still ultimately dependent on price, Piednoël said. Cost is therefore critical to the decision to go with a single chiplet tape-out, especially as a startup up against the aggressive pricing strategies of competitors.

“Our competitors can squeeze really hard on price,” he said. “There is no way to win on pricing with multiple tape-outs.”

Chiplet-based disaggregated compute efforts in Europe and Japan that require multiple chiplet tape-outs will simply work out to be too expensive, Piednoël said.

“Both the market leaders in physical AI are talking about chiplets as a distraction—they do monolithic and they do it very well,” he said. “We have to find something we do better than them, which is already a tall order, while being able to engage on pricing. The solution was to come down to one single tape-out, without advanced packaging.”

Having full control over the design of Athos’ chiplet is also a big plus.

“When you put chiplets together on a substrate, you can’t just put any chiplets together,” Piednoël said. “You need to make sure that they are balanced so that one isn’t dumping a constant current into the other one. This is more complicated than it looks, and not just at power-on, while it’s running, too.”

Controlling its own chiplet manufacturing means chiplets can be binned such that those with the lowest leakage are kept for applications with the most stringent safety requirements, Piednoël said. Combining third-party chiplets also brings challenges around matching capacity and impedance of connections in a reliable way, especially given automotive’s 15-year lifetime requirements.

In safety-critical systems, dual power sources are often required for redundancy. Athos’ design has redundant power rails and dual voltage regulators for each chiplet. While the company uses well-known schemes to avoid cascading, such as reducing the clock frequency above a certain temperature, a custom voltage regulator design from a European manufacturer will add to safety. An eFuse inside each regulator measures current and shares it over I2C for neighboring chiplets to decide whether there’s a problem or not.

In the new design, Athos’s identical compute chiplets, each an SoC, will be able to vote using the same mechanism as the earlier design, which is one of Athos’s key patents.

In this scheme, each chiplet monitors its neighbors for software and hardware issues via information on scheduling, temperature, clocks, and other health metrics. Redundancy across more than two chiplets means the chiplets can vote, deciding between themselves which chiplet is having issues and needs to be reset. “Failed” chiplets become available to back up the others once they are back online.

“Most of the industry uses lockstep, but lockstep doesn’t tell you which one is wrong; it just tells you there’s a problem,” Piednoël said. “Here, we have full disambiguation of the errors. The information means you can immediately diagnose the problem, replace the chiplet, and have no discontinuity in computing.”

Most of the time, it’s the memory that causes failures when high-energy particles hit, causing bit flips.

“If the processor reads the value, which is wrong, maybe it has a division by zero or an exception, so your chiplet misbehaves,” Piednoël said. “You’d never figure that out with lockstep.”

In another common scenario, a software thread is supposed to finish but doesn’t; another chiplet can take over until the original one can be reset.

“This is bulletproof against division by zero,” Piednoël said. “It can even protect against hackers that try to do overflow buffers, because we check all of our buffers.”

In another easy-to-imagine scenario, the car is driving with Level 3 autonomy when the driver appears to fall asleep, such that the car needs to switch to Level 4. This is a tricky scenario for other solutions, Piednoël said, because you can’t stop running Level 3 driving while you switch to Level 4. With Athos’ design, more chiplets can be brought online to run the Level 4 during the transition. This means fewer backup chiplets available during the transition, but this risk is acceptable for a few seconds, Piednoël said.

One or more of Athos’ Android-compatible chiplets could also be used to run non-driving functions like infotainment, which can quickly be switched off if the chiplet is required for safety reasons. Switching from infotainment to driving tasks would take as little as 10-15 milliseconds, Piednoël said.

Using multiple identical compute chiplets also gives Athos the opportunity to run built-in self-test programs like MBIST and LBIST on chiplets that are idle, to identify problems before they happen.

“While you’re driving, you need to be doing some tests on your chiplets,” Piednoël said. “While MBIST or LBIST can pass while you’re parked in your garage, it may not pass while you’re driving through the Mojave Desert, and the water cooling everything is hotter.”

Many autonomous driving accidents happen due to long-tail events, so Athos has also developed IP around trajectory choice.

“Machine learning usually gives you around 20 possible trajectories, and the car chooses which one to take, based on comfort,” Piednoël said. “If we have any doubt that it is not free space, we say that [the] trajectory is not good, give me another one.”

This idea uses an algorithm developed by Mercedes-Benz, which relies on laser and radar sensor data to raise the certainty of predictions about whether a particular area of space around a car is empty or not. Unlike trajectory planning algorithms, it doesn’t use machine learning, but combining the two to rank possible trajectories in order of safety is the safest possible idea, Piednoël said.

The ultimate goal is a system that’s statistically one million times safer than the minimum required by ISO 26262. When considering pricing, OEMs should also keep in mind their liability for accidents, Piednoël said.

“The level of deterministic behaviour we have is what’s going to save your butt in court later,” he said.

Athos will work on two key markets initially: autonomous driving and drones. Automotive’s long design cycle time means the startup needed a secondary market that can generate revenue in the shorter term. Drones are ideal because they require powerful AI capabilities with a lower functional safety requirement, which means they can be served by a single-chiplet version of Athos’ design.

“We will add a zero to the number of TOPS compared to what’s available on the market, around 250 TOPS,” Piednoël said.

Scaling up or down, and custom versions of the design, are easy, given the design’s relatively cheap organic substrate, Piednoël said. Other markets could be manufacturing and robotics, where significant intelligence is required, combined with functional safety.