Interview with Bart Placklé, Imec
“We need leading edge solutions, which form the heart of the car”
Bart Placklé, Vice President Automotive at Imec.
Imec
Chiplet architectures are redefining the automotive industry. Imec’s Bart Placklé explains how modular design, interoperability, and ecosystem collaboration can help Europe regain leadership in high-performance vehicle computing.
With more than two decades of experience in automotive and
semiconductor technology, Bart Placklé has shaped some of the most fundamental
transformations in the industry – from infotainment consolidation and domain
controller evolution to Intel’s early autonomous
mobility programs. As Vice President Automotive at Imec, he now leads
efforts to advance chiplet integration as a foundation for next-generation
high-performance automotive computing.
At the Automotive Computing
Conference 2025, Placklé will speak about the structural and economic
limits of traditional SoC design and why open, interoperable chiplet ecosystems
are the key to scalable, sovereign computing in Europe. In the run-up to the
conference, we spoke with him about the technological and structural shifts
driving the future of automotive computing.
ADT: Automotive computing performance is reaching
physical and economic limits in traditional System-on-Chip design. From your
perspective, how can chiplets help the industry regain leadership in
high-performance automotive computing?
Placklé: By cars becoming semi-autonomous, the
overall compute requirements for self-driving artificial intelligence and
inference are exponentially growing (the perception and prediction of the
environment around the car). At the same time, thanks to automated driving, the
car really becomes a true, perfect entertainment or office on wheels again,
driving the in-cabin compute requirements to unseen levels through the need for
edge agentic artificial intelligence. If you now combine these challenges with
the fact that most high-performance electronic control units are consolidating
into a single high-performance computing control unit, it is clear that the
compute requirements continue to surge beyond anything we have ever seen. But
the challenge does not stop here.
How do these escalating performance demands translate
into economic and structural challenges for the industry?
Even if automated driving and advanced driver assistance
systems are installed in most of the roughly 90 million cars produced each
year—a figure that continues to decline—the automotive industry still operates
at volumes far below those of smartphones, yet demands computing performance
that is vastly higher. So, from a business-case perspective, we have a
two-orders-of-magnitude problem, with the latest monolithic advanced solutions
already costing around two billion to develop, resulting in only a few players
remaining. I think the statement "this is not on a good trajectory"
is an understatement. Everything I mentioned was written in the stars and is
why we started taking a right-hand turn by building such high-performance
computing systems differently. And if you do not think this was challenging
enough, we did not anticipate the geopolitical climate, where the supply chain
is no longer a given. The heart and DNA of the car are software and
high-performance computing, and this is where Europe has a real problem: we do
not have leading edge solutions, which form the heart of the car. That has to
change, and it has to change now.
So where do you see the starting point for this change?
How can Europe realistically catch up in high-performance computing and
semiconductor technology?
We do not have leading edge solutions, but we have very good
building blocks, as seen in the stellar European example where Airbus was able
to build the most successful airplanes not by trying to do everything
themselves but by pulling all state-of-the-art assets from the region. That is
exactly what we are doing with automotive chiplets. We have best-in-class
artificial intelligence accelerators, high-performance CPU architectures like
Arm and later RISC-V. We have companies like NXP and Infineon that are excelling
in input/output, safety, and cybersecurity, and Bosch is creating the
automotive context with a base die. We have all the right technology to pull
this together. Building the "Airbus" for automotive high-performance
computing is the path to a state-of-the-art automotive sovereignty (robotics
will adopt the same). To make this ready and economically scalable, we need to
ensure the chiplets are interoperable (can be developed independently of each
other and "late bound," assembled on a package) and that the
technology survives the strong automotive environmental mission profiles. By
aligning the ecosystem through interoperability and focused research on quality
and reliability, we aim to accelerate and de-risk the adoption of automotive
chiplets — paving the way toward sovereign, state-of-the-art vehicle
technology.
At the ACC 2025, you will
discuss interface standardization as a path to cost-competitive chiplet
solutions. What progress has been made toward interoperability, and what still
needs to happen for true scalability across suppliers?
When we started, the first focus was convincing the industry
that the right-hand turn to chiplets is the right way to go. Today we do not
have to convince anyone about chiplets, but to your question, what is needed
for chiplets to "fly" (or drive) is seamless chiplet
interoperability. If single vertical ecosystems stick to their proprietary set
of chiplets, you will not have the scale and re-use across different suppliers.
You end up with a cost structure that could be even higher than the reticle-size
monolithic solutions, as the heterogeneous and yield savings will not be enough
to offset the need for doing all the chiplet tape-outs yourself. The whole
value only works when you have multiple CPU, GPU, base die, and artificial
intelligence units that can be used in different constellations for different
performance levels and customers. That is key to affordability.
What needs to be in place to make such scalable chiplet
solutions truly interoperable across different vendors and architectures?
The chiplet success caused a whirlwind in the ecosystem,
where everyone now starts having a chiplet product, but they are mainly
compatible only with themselves. Our first focus right now is to create a
reference specification to ensure chiplets can talk to each other. To really
allow late binding, it is by far not sufficient that the die-to-die physical
layer is aligned. The Universal Chiplet Interconnect Express does a great job
on the physical layer, but late-binding interoperability requires alignment on
the higher-level protocols, sidebands, boot, safety, and software constructs.
With the Imec Automotive Chiplet Platform Reference Architecture Specification,
we are trying to glue all these dependencies together—not with the goal of
reinventing the wheel, but to pull all relevant standards together and steer
existing standards where needed to achieve end-to-end chiplet interoperability.
Imec has been instrumental in driving chiplet innovation
from concept to prototype. How do you see collaboration between research
institutes, foundries, and OEMs evolving to turn chiplet architectures into a
real automotive success story?
Our Imec automotive chiplet program is now entering its
third year. We modeled various die-to-die interoperability levels and built
several thermal-mechanical test vehicles to understand how far we can take
packaging technology that fulfills the automotive cost, performance, and
quality requirements. While this has been a great success, we need to move fast
and bring the learning closer to production, as we do not have five years to
make this turn. Therefore, we decided to go one step further, we no longer just
focus on thermal-mechanical vehicles. In Heilbronn (a city in southern Germany
– editor’s note), we will start building functional reference designs by
combining existing and future chiplets from the ecosystem as a way to
accelerate and de-risk production.
What additional insights do you expect from actually
building and testing these reference designs, compared to purely simulated or
theoretical models?
By physically building such chiplets on a package and
bringing them up, there will be many learnings we can share with the
ecosystem—from supply chain challenges when wafers and chiplets are being
passed around to implementing safety, boot, power, cooling, and system
management. Everyone needs to go through the learning curve when they move to
chiplets. Every time we integrate a new chiplet, we will encounter new
challenges, such as small package or protocol mismatches—and who knows what
else we have not thought of yet. Heilbronn is going to be a test bed for
chiplet integration, just as core Imec is a test bed for new process
technologies—to learn what fails, recommend what works, and pursue one goal:
accelerating and de-risking the adoption of automotive chiplets.
