Ruhr-based Gemesys, which is designing chips that “mimic the behaviour of the human brain”, has raised €8.6m in pre-seed funding. It plans to use the money to more than double its headcount and for R&D efforts.
Atlantic Labs and Amadeus Apex Technology Fund led the round and were joined by NRW.Bank, Sony Innovation Fund and Plug and Play. The round also included a grant from the German government through the country’s EXIST ‘transfer of research’ programme, which backs research-based companies. CEO Dennis Michaelis declined to comment on the split between equity and grant funding.
What does Gemesys do?
Gemesys is designing chips that allow AI models to be trained and run directly on devices like smartphones and hearing implants. This eliminates the need to transmit any raw data to a data centre and reduces the energy use associated with traditional model training.
Central to the company’s technology is the use of memristors — electrical components that allow artificial neural networks to resemble biological synaptic functions (the way the human brain processes and acts on information) — on its chips.
Memristors give the potential to develop a “neuro-inspired type of computer”, Michaelis says. “The human brain is the best processor for intelligence that we know, a million times more efficient than any AI chip on the market,” he adds.
A brain-inspired chip has the potential to get around a key bottleneck in the training of AI, Michaelis tells Sifted. According to him, current chip architecture is limited because where the data is stored and where it’s processed on the chip are separated. “Whenever you want to do a computation you have to read the data from the memory, load it into the processor, compute and then write the result back. The problem is you have to shift data along at every step,” he says. This data shifting slows down the response time and uses more energy than is needed to complete a task.
“Such a bottleneck doesn’t exist in the human brain,” Michaelis says. “It’s not separated into two units […] it all happens in the same place.”
Gemesys’s chips could have two main benefits. One is that by reducing the energy consumption associated with AI by doing most of the work on the device, they reduce the role of data centres. Research by Goldman Sachs predicts that power demand in data centres because of the AI boom will increase by 160% by 2030 — meaning data centres will contribute 3-4% of global power demand, up from 1-2% today.
The other benefit is that they aim to give users better data privacy. By doing the training and running the model on the device, you don’t need to send any raw data to the cloud.
“You can design models that are a lot smaller and have fewer parameters to do the same task,” Michaelis says. “We’re aiming for 10k times less energy use.”
The chip architecture is based on research Michaelis did at Germany’s Ruhr University Bochum with one of the company’s cofounders and chief scientific officer Enver Solan. The company’s other cofounders are CFO Moritz Schmidt and CTO Daniel Krüger.
The company counts Dr Christian Wenger, professor for semiconductor materials at Brandenburg University of Technology, and Jamie Urquhart, cofounder and former COO of Arm, as advisors. Its board is made up of Quentin Calleja from Atlantic Labs, Ion Hauer from Apex Ventures and Massi Ali Ahmadi from NRW.Bank.
How will it make money?
The company plans to go to market with a business model similar to chip giant Arm. It’ll design its chips in-house and then outsource the manufacturing to semiconductor manufacturers like TSMC. It wants to sell to original equipment manufacturers like Bosch, Apple and Siemens, which it sees using its chips in their products.
Gemesys is currently working with three pilot customers — one in the energy grid sector, one that works on cochlear implants (electronic devices that help people who are deaf or hard-of-hearing to hear) and another that does security. Michaelis says use cases extend to any company working on IoT devices.
What’s next?
Gemesys is planning to more than double headcount in the next few years, from 19 staff. It has seven open positions across its hardware development, concept and research teams. Hiring will be the company’s “biggest expense” in the next couple of years, Michaelis says.
Attracting talent can be a problem in Europe, especially for companies that want the best people working on chip hardware. They tend to be concentrated around the biggest chip manufacturing plants, which are currently in North America and Asia, Michaelis says.
The company mainly needs to hire in its software team though, which Michaelis says shouldn’t be a problem. “We have a lot of great computer scientists in Germany, and in Europe,” he says. “That’s really good.”
