Synthara: Shaping the Future of Edge Intelligence
Edge devices like wearables or IoT devices produce a lot of data, and processing this data is key to delivering better services. To avoid sending all the data back and forth to the cloud, one aims to process it on the edge device directly.
However, unlike cloud servers, edge devices have limited energy resources, so improving energy efficiency and avoiding head generation is a major challenge. That’s where custom chip architectures come to the rescue.
Spun out by Manu Nair and Alessandro Aimar from the Institute of Neuroinformatics of the University of Zurich and ETH Zurich, the semiconductor startup Synthara develops next-level AI chips for the edge—combining high performance with energy efficiency. Since their oversubscribed seed round in December 2020, led by High-Tech Gründerfonds and with the participation of the Zürcher Kantonalbank and several business angels. They have started their early adopter program and made progress towards launching their first products.
Learn more about the future of edge intelligence from our interview with the CEO, Manu Nair:
Why Did You Start Synthara?
I think the purpose of a company is to find the best possible way of solving a problem. And the problem I was excited about was building the most innovative chip for data processing.
My co-founder Alessandro and I met during our studies, and we both came up with a long-term plan to found an AI chip startup. That’s why we decided to start our PhDs at the Institute of Neuroinformatics of the University of Zurich and ETH Zurich and conduct all the research that’s now at the core of Synthara.
How Does Designing AI Chips Work?
Computing today involves a lot of data being shuffled around between CPU and memory, which consumes a lot of energy and produces heat. As one of my professors liked to say: It is like going from the train station—the CPU—to the warehouse – the memory – every time you want to make a computation. But what if you could make computations directly inside the warehouse?
We were rethinking chip design from first principles: From physics laws that describe electron flow and heat dissipation. Through lots of thinking and testing, we iterated chip designs quickly, trying to identify flaws as early as possible. We came up with a new chip design philosophy that eliminates most of the data flow through so-called in-memory computing and is geared toward edge applications.
So, on the one hand, we now have software that takes the constraints of the use case as input and configures the chip design within our design philosophy for that particular use case – while still being flexible enough to support adjacent use cases.
On the other hand, our neural network accelerator Adaptiva delivers up to 100 terra operations per second per watt of energy efficiency—that’s trillions of operations. And it could go pretty crazy, up to 1000 terra operations per second per watt. The ceiling is only physics: noise, and the fact that some data still needs to move around.
How Did You Evaluate Your Startup Idea?
There are several metrics to benchmark our technology: On the one hand, energy efficiency in terra operations per second per watt. On the other hand, latency – as we want networks to respond fast, e.g., for wearables, the chip performance per area, or the maximum frequency and throughput of the chip.
We excel in energy efficiency. We tested this by running neural networks on in-memory chips, which you can read about, e.g., in this arXiv paper. We demonstrated at least 500× higher energy efficiency using our models on compatible neuromorphic chips compared to Cortex-M4, a popular embedded microprocessor.
What Advice Would You Give Fellow Deep Tech Founders?
Start talking to potential customers as early as possible! Figuring out how to talk to customers is the hardest part, and the sooner you understand your customer, the better. Most founders know their tech quite well. What trips them up are all the peripheral but non-trivial reasons why people cannot use it.
At Synthara, we’re still in the process of this. We launched our early adopter program, which proved invaluable for quickly and reliably obtaining early customer feedback.