back

»Our goal is to develop a chip design that does not exist in this form today and that adds significant value to the AI-optimized hardware industry«

The Bochum-based start-up GEMESYS Technologies emerged as the winner of the KUER.NRW Business Plan Competition 2022, beating more than 30 start-up teams from the future-oriented industries of climate, environment, energy efficiency and resource conservation (KUER). The competition is carried out within the context of KUER.NRW on behalf of the Ministry of the Environment, Nature and Transport of the State of North Rhine-Westphalia. The GEMESYS team is working on recreating the information processing of the human brain in electrical circuits. In the brain, there is no separation between the information-processing and information-storing unit – unlike in conventional computer architecture, where data is continuously exchanged between processor (CPU) and memory (RAM). A process that is energy-intensive and comparatively slow. This is precisely where the GEMESYS team comes in. In an interview with KI.NRW, GEMESYS Managing Director Dr.-Ing. Dennis Michaelis outlined the paradigm shift at stake in the development of memristors.

You set out to lay the foundation for a chip design »made in NRW« – specifically, of computer chips which don’t even exist in this form today. How did that come about?
The idea to develop a chip design »made in NRW« originated from the dissertation writings of my co-founder Dr.-Ing. Enver Solan and me. During our time as PhD students, we shared an office and in many of our discussions we noticed that there is a need in the AI-optimized hardware industry for innovative and powerful concepts that meet the current requirements and challenges. Together with my second co-founder Moritz Schmidt, we created GEMESYS to fill this gap in the market and to develop a new chip design that meets the requirements of the future.

We rely on an interdisciplinary approach, combining expertise from different fields such as electrical engineering, materials science, and computer science, and in the future also areas such as neuroscience, psychology, and ethics. We work closely with partners from research and industry to drive the development process and ensure that our chip designs are also practically feasible and economically viable.

Our goal is to develop a chip design that does not exist in this form today and that offers significant added value for the AI-optimized hardware industry. We believe that our work will help to establish NRW as a leading location for the development of AI-optimized hardware solutions, while also opening up global markets.

So-called memristors are a key part of your technology: What is new about a memristor and what are its advantages, particularly regarding artificial intelligence?
A memristor is an electronic component that can change electrical resistance and thus regulate the flow of current. What makes a memristor special compared to conventional electronic components such as transistors is that it can store not only the resistance value, but also the state it was last in. This makes it possible to process and store information in the same physical location – creating what is known as in-memory computing.

In terms of artificial intelligence, memristors offer a number of advantages. They can process large amounts of data quickly and efficiently, which is particularly important for complex machine learning applications. They also enable neural networks to be implemented directly at hardware level, which leads to a significant reduction in energy consumption. Memristors have the potential to realize neural networks with billions of parameters on the surface of a fingernail.

You have been able to prove the technical feasibility of your chip design in software emulations. Can you describe an emulation of this kind?
A software emulation is a process in which the functions and properties of a computer chip are emulated in a virtual environment. In this process, the electrical signals that the chip would normally process are emulated by software. This means that the chip does not have to be physically present in order to test or optimize its functionality.

For this purpose, we wrote our own emulation environment, which is specifically tailored to our requirements and needs and allows us to test and optimize the behavior of our chip in different scenarios.

After having ensured the conceptual functionality, we now move on to hardware-related investigations, where we can, for instance, assess the exact power consumption or the temperature distribution. Various error sources and failure causes can also be investigated to improve the robustness of the design.

The results of the software emulation are used to further optimize and refine the chip design. Only when the design has been fully tested and validated in emulation does it go into production to manufacture a physical chip. Overall, software emulation provides an efficient and cost-effective way to develop and test a chip design before it is physically manufactured.

So now concrete use cases could be tested together with innovative industry partners. Artificial intelligence also comes into play here. What specific collaborations and areas of application are you thinking of?
We expect to perform processing with AI to calculate large correlations faster and more energy-efficiently. This is sorely needed, as DeepMind’s Alpha Go cost about $35 million to train, Tesla’s Full Self-Driving AI on its custom Dojo chip takes over a month to train, and OpenAI’s ChatGPT-3 consumes millions of dollars every week to run.

Powerful AI is also being used in other fields such as drug development or nuclear fusion, two very exciting fields.

Our technology also allows AI to be moved from the cloud directly to the places where the data is generated, the so-called edge of the network. That’s why this field is called edge AI. Training and applying AI directly on site comes with some significant advantages: increased data protection, lower latency, lower energy consumption, and possible offline operation. Interesting application areas arise where one or more of these characteristics still prevent AI from being implemented with current state of the art. Because this is where GEMESYS can not only improve something that already exists but create something fundamentally new.

There’s a lot going on in hardware research – and internationally, too. Is it possible to roll up global markets as a start-up and from NRW?
Yes, it is certainly possible for a start-up from NRW to open up global markets in the research field of AI-optimized hardware. There are many successful examples of start-ups that started out in the region and are now operating worldwide: Lidrotec uses lasers to cut microchips in Bochum and is increasingly active in North America, and eleQtron based in Siegen is competing with companies such as Google in the field of quantum computers. Some of the factors that contribute to this are, for example, a globally competitive business strategy, an innovative product, an experienced team and a strong financing base.

In NRW, there are a large number of institutions supporting start-ups, including in the area of AI-optimized hardware. There are funding programs, consulting services and networks that support start-ups in developing and marketing their products. This is why the work of KI.NRW is so important. In addition to the excellent work of the start-up Center at Ruhr University Bochum, the WorldFactory, for example, High-Tech.NRW’s accelerator program has helped us a great deal.