Software Optimises Chip Design for Energy-Efficient Artificial Intelligence Applications.

SENMap, a new software tool, optimises the mapping of large spiking and artificial neural networks onto adaptable hardware architectures. Evaluations demonstrate a 40% reduction in energy consumption compared to a baseline system, facilitating efficient chip design prior to fabrication and enabling future network modifications.

The increasing demand for artificial intelligence at the network edge necessitates computational paradigms beyond conventional von Neumann architectures. Neuromorphic computing, inspired by the efficiency of the brain, offers a potential solution, but realising its benefits requires sophisticated tools to translate algorithms into hardware configurations. Researchers are now detailing SENMap, a software framework designed to optimise the mapping of both spiking and artificial neural networks onto adaptable neuromorphic systems. This work, presented in a paper by Prithvish V Nembhani (University of Manchester), Oliver Rhodes (University of Manchester), Guangzhi Tang (Maastricht University), Alexandra F Dobrita (imec-NL), Yingfu Xu (imec-NL), Kanishkan Vadivel (imec-NL), Kevin Shidqi (imec-NL), Paul Detterer (imec-NL), Mario Konijnenburg (imec-NL), Gert-Jan van Schaik (imec-NL), Manolis Sifalakis (imec-NL), Zaid Al-Ars (Delft University of Technology) and Amirreza Yousefzadeh (University of Twente), introduces a flexible mapping strategy that demonstrably improves energy efficiency – achieving a 40% reduction in a baseline configuration – while accommodating the complexities of large-scale neural network deployment. The team’s work, titled ‘SENMap: Multi-objective data-flow mapping and synthesis for hybrid scalable neuromorphic systems’, is supported by the open-source emulator SENSIM, facilitating chip design prior to fabrication.

Optimised Mapping Strategies Advance Neuromorphic Computing

Recent research focuses on translating spiking neural networks (SNNs) into practical implementations on neuromorphic hardware, achieving substantial reductions in energy consumption. Investigations consistently address the challenge of efficiently allocating SNN computations onto specific architectures, such as three-dimensional Network-on-Chip (3D-NoC) systems, with a clear emphasis on optimising for energy consumption and throughput. Current efforts employ established optimisation algorithms, including Particle Swarm Optimisation (PSO) and parallel computing frameworks like Dask, to navigate the complexities of large-scale SNN mapping.

Spiking neural networks (SNNs) represent a departure from traditional artificial neural networks (ANNs). ANNs process continuous values, whereas SNNs utilise discrete ‘spikes’ – brief electrical pulses – mimicking neuronal activity in the brain. This event-driven computation promises significant energy efficiency gains, particularly when implemented on dedicated neuromorphic hardware. Neuromorphic chips aim to replicate the structure and function of the brain, potentially offering advantages in power consumption and processing speed for specific tasks, but effectively mapping SNNs onto these architectures presents considerable challenges.

Network-on-Chip (NoC) systems represent a common architecture for on-chip communication, and three-dimensional (3D) implementations offer increased bandwidth and reduced latency. Optimising SNNs for 3D-NoC systems demands careful consideration of data routing and resource allocation, requiring sophisticated algorithms for the allocation of computational resources, scheduling of spikes, and optimisation of network parameters.

SENMap and SENSIM represent tools developed to address these challenges. SENMap functions as a mapping algorithm that translates SNNs and ANNs onto adaptable hardware architectures. SENSIM operates as an emulator, simulating the behaviour of mapped networks to evaluate performance metrics before hardware implementation. The open-source nature of these tools fosters collaboration and accelerates innovation in the field of neuromorphic computing, providing a flexible foundation for future advancements. Researchers can explore architectural modifications and mapping strategies.

SENMap demonstrates effectiveness, improving energy efficiency by up to 40 per cent compared to a baseline asynchronous implementation as network size grows. The open-source nature of both SENMap and SENSIM promotes collaborative development and accelerates the design process for neuromorphic chips, providing accessibility for researchers to explore architectural modifications and mapping strategies before committing to fabrication.

Researchers actively explore automated design space exploration techniques, coupled with co-design approaches that optimise both software and hardware simultaneously, refining mapping algorithms to accommodate increasingly complex network topologies and diverse hardware constraints. Investigations into the resilience of mapped SNNs to hardware variations and the development of techniques for dynamic reconfiguration represent important areas of focus, ensuring robust and adaptable neuromorphic systems. Future work will likely concentrate on these areas, pushing the boundaries of neuromorphic computing and unlocking its full potential.

The development of hybrid SNN-ANN approaches indicates an attempt to leverage the strengths of both paradigms for improved performance and efficiency, combining the biological realism of SNNs with the established performance of ANNs.