SEMRON: Shaping the Future of AI Chips

With the advent of deep learning, application-specific chip architectures have opened up a new market opportunity. Especially running neural networks for inference on edge and IoT devices requires ultra-low-power, custom-built AI chips that depart from the general-purpose chip architectures dominating the last two decades. 

One approach pursued by major corporates and several startups is to design application-specific integrated circuits (ASICs) based on new silicon chip designs. However, much larger performance improvements may be possible by rethinking chip design from first principles and asking: What physics principles can we employ to build the most energy-efficient AI chip?  

Aron Kirschen and Kai-Uwe Demasius have been through this exercise in their studies and research and eventually founded the startup SEMRON in February  2020. By using capacitive coupling, their AI chip avoids Ohmic losses and minimizes heat dissipation, which allows for stacking several layers on top of each other – and enabling truly three-dimensional chip architectures while still being CMOS-compatible. After mastering the Herculean endeavor to get fab capacity in a period of chip shortage and receiving the proofs from the fab, SEMRON is now preparing to raise its next round.  

Learn more about the future of AI chips in our interview with the CEO Aron Kirschen:

Why Did You Start Semron?

I met my now co-founder Kai in my first term at university, and we both realized that machine learning is impaired by the end of Moore`s law. We had to find a new paradigm for the whole computing architecture, and around the same time, some visionaries were talking about analog in-memory computing. We thought this would be going big  – and that we wanted to be part of it. Everyone wanted to be a TensorFlow expert. But software can only build on what hardware is available – and while software was developing at breakneck speed, hardware geared towards AI applications lagged behind. 

We developed the idea for SEMRON after we went on a study trip to Cambridge about ten years ago. Machine learning software consumes so much energy, whereas human brains need, in comparison, so little – how could we get closer to brain-like energy efficiency? We tried to envision what that would mean on a chip level – and geared all of our studies and research towards building the most energy-efficient AI chip possible. While Kai did a Ph.D. to learn more about semiconductor technology, I also studied maths since you will find it in every modern technology (and simply because it is incredibly beautiful). 

I want to see as much as possible of the future – that’s why I am working on the future myself. Working on something that could become like a human brain was scary, but AI was here to stay, and we wanted to be part of it.

How Does Your AI Chip Work?

While general-purpose chips run artificial neural networks through bit operations, neuromorphic chips emulate them by hardware design: Imagine you trained a neural network, and now you want to run it on an edge device with very low power consumption.  The neuromorphic chip then encodes the weights and biases of that particular artificial neural network in an analog manner to run it very efficiently for inferences. 

One known example are memristive devices that encode the weights of neurons by the electrical resistance of memristors and use the fact that electrical currents add up to integrate the signal from several neurons. However, electrical resistance always leads to heat generation – that’s why we designed our AI chips based on memcapacitors: The weights of neurons are encoded by the capacity of memcapacitors, and the signal from several neurons is integrated by the charge accumulation from all the memcapacitors. 

For computing hardware, there are different layers of innovation: Software innovation, circuit innovation, device architecture innovation, and material innovation. We are innovating on the level of the device architecture, but we are still CMOS compatible.

And talking about energy efficiency: Did you know that penguins are taller in the Antarctica? As their volume-to-surface ratio increases, they keep more heat inside. With chips, it is the exact opposite: We want to dissipate the generated heat as much as we can. The more silicon layers we stack, the more energy we have to lose. However, the surface remains more or less the same. That’s why processors so far have been two-dimensional. Only because our chips are so energy efficient, i.e. generate so little heat, can we stack several layers on top of each other and build truly three-dimensional AI chips. 

How Did You Evaluate Your Startup Idea?

This question is deeply ingrained into our company culture, especially as we attempt to create something fundamentally new: How do you prevent wasting parts of your life because you are lying to yourself about the potential of your technology?

We are solving this through our company culture – we are our own greatest critics. Everyone is allowed to say what they think. We challenge our approach constantly and don’t fear treading on someone’s toes. We expect everyone to continue asking until the right abstraction level is reached. If someone is explaining the details first, please interrupt as fast as you can and don’t accept the explanation – because it isn’t one. Being ‘impolite’ that way is part of our company culture. And our employees should use Semron to increase their learning curve. If their learning plateaus, we would actively support them to land a better-suited position at a different company. 

Also, we are very open with our innovation and validation – e.g. we published a Nature paper outlining the fundamentals of our technology – which also served as external validation. And we are rigorously benchmarking our technology – just recently, we got the first fab-out silicon implementing our design. Being silicon-backed is a huge step for our technology. 

Company culture as a hack: Always ask yourself whether you’re on track. You need passion to work on deep tech, as it may take a few years longer to realize gains. Be transparent with tech investors about the potential of your technology. There’s no point in overselling.