Interview with Harald Kroeger, SiMa.ai
“Putting the ghost in the machine”
Harald Kroeger holds an M.S. in Electrical Engineering from Stanford University and a Bachelor’s degree in Economics from the University of Hannover.
Martin Stollberg
As AI moves from the cloud into the vehicle, compute architectures must deliver low latency, efficiency and scalability. Harald Kroeger of SiMa.ai explains why edge-first AI will shape the next phase of software-defined vehicles.
AI is becoming one of the defining
forces behind the next generation of vehicle architectures, but scaling
it inside the car requires more than additional compute performance. For
software-defined vehicles, the decisive questions increasingly concern latency, energy efficiency, local inferencing and the right
balance between cloud and edge intelligence.
Harald Kroeger, Head of Sales & President Automotive at
SiMa.ai, brings long-standing automotive, semiconductor and systems experience
to this discussion. At the Automobil-Elektronik
Kongress in Ludwigsburg, he will join the panel discussion “Chiplets –
Technology & Business Viability”. In this interview, Kroeger explains why
AI-defined vehicles may become the next step beyond SDVs, how AI accelerators
are changing vehicle compute architectures and where current hardware
strategies still fall short.
ADT: Looking ahead three to five years, what will be the
biggest bottleneck in turning SDV and AI strategies into scalable,
industrialised vehicle platforms?
The AI-defined vehicle will eat the SDV for lunch – it is
important to take the next step now. The solution will be edge-first –
cloud-first solutions will suffer from latency and high running costs. It is
paramount to architect these two facts into future vehicles now.
Which decision being made today will most strongly
determine where value is created in the future automotive ecosystem?
Having enough local AI inferencing
power available in the vehicle is critical – automotive platforms today
need to provide enough headroom for this. Additionally, choosing silicon that
is backed by a strong software platform is essential; otherwise, it risks
becoming an underutilised asset.
Where do current approaches to SDVs and next-generation
E/E architectures still fall short in real-world programmes?
There is too much dependency on Swiss Army knife types of
silicon: the future will have chips specialised for different tasks.
How are AI accelerators changing vehicle compute architectures?
Adding AI firepower is crucial – this will be outside the
usual all-rounder chips for IVI or ADAS, as this is the more scalable and
cost-efficient architecture.
What are the biggest challenges in scaling AI workloads
in automotive systems?
The automotive industry is just
beginning to grasp the possibilities of GenAI in the car
– on the edge. SiMa.ai calls this putting “the ghost in the machine".
Overemphasizing the cloud not only increases cloud and compute costs but also
causes latency delays and a less natural user experience, which can frustrate
users.
Where do you see the biggest mismatch between AI
expectations and hardware capabilities today?
OEMs and Tier 1s are traditionally slow in adopting
new silicon – to win the race, companies need to anticipate the future and bet
on virtual silicon, meaning silicon that is
fully defined and can be emulated, but does not yet exist as physical silicon.
