Neural pathways inspire computing breakthroughs
In order to serve a diverse range of artificial intelligence (AI) applications, electronic devices are becoming increasingly smart. This requires embedding ever more intelligence and computing power, while keeping these devices compact and energy-efficient. “While different technology options have been explored in the past, none have been found to have the approximate processing efficiency of the human brain,” explains TEMPO project coordinator Björn Debaillie from imec in Belgium. “In this project, we wanted to move away from classical computing architecture, and mimic the structure and functions of the human brain in semiconductor hardware.”
Neuromorphic hardware technologies
The aim of the TEMPO project was therefore to develop neuromorphic hardware technologies, based on the structures, processes and capacities of biological brains. The project team also wanted to broaden the deployment of neuromorphic hardware over different applications. “Disruptive hardware solutions are essential to sustain the migration of AI processing from the cloud to the edge for improved performance, service, reliability and safety,” says Debaillie. “To achieve this, we developed benchmarks and roadmaps to support the assessment of emerging technologies. We also looked at optimising European infrastructure for the manufacturing of neuromorphic chips.” To ensure industry-viable solutions, key industry players across the entire value chain were involved in the project. Open collaboration within the consortium was established, to leverage cutting-edge expertise as well as semiconductor research and manufacturing infrastructures. “One focus of our research was on enhancing emerging memory technologies, to achieve powerful yet efficient data processing,” notes Debaillie. “We also explored the entire neuromorphic technology landscape, covering a wide range of production-ready maturity levels.” Different algorithms were benchmarked, and their potential applicability evaluated and validated over four key application domains: smart health, smart electronics, smart industries and smart mobility.
Harmonised semiconductor fabrication
This work has helped to lay the foundation for realising an AI shift from the cloud to smart devices. The project’s achievements have been published in more than 100 scientific articles, and resulted in the development of 11 proof of concept demonstrations across the four target application domains. Key breakthroughs have also been achieved in the optimisation of emerging memory technologies, ready for implementation in neuromorphic hardware. “Comparisons between different technologies, architectures and approaches have been captured in a holistic benchmark,” says Debaillie. “This will enable educated technology choices along the entire development cycle and value chain.” The TEMPO project team also established mechanisms to harmonise semiconductor fabrication. “This will enable the building of centres of excellence in terms of expertise and infrastructure, enhance complementing fabrication, and strengthen Europe’s semiconductor offering,” adds Debaillie.
Neuromorphic hardware technologies’ benefits
The neuromorphic hardware technologies pioneered in TEMPO could bring everyday benefits. “Smart devices will offer more applications and services, with better performance, security and safety, because data can be processed locally in the device,” remarks Debaillie. TEMPO’s work has been continued through collaborative research projects such as ANDANTE and been taken up by innovative SMEs and industrial pioneers. Through covering the entire value chain, from technologies to applications, the project has played a key role in strengthening Europe’s position in cutting-edge computing. Indeed, the results are also likely to encourage further research activities. “Market adoption of these technologies is evolving gradually,” observes Debaillie. “This trend is clearly visible, as the term Edge AI is gaining attention and ‘neuromorphic computing’ is considered the key to enable its implementation.”
Keywords
TEMPO, computing, biological, brains, neuromorphic, semiconductor, algorithms
