What is it about?
In this work, we introduce NEUROPULS, a Horizon European research project focused on the development of a secure, energy-efficient accelerator powered by integrated photonics for edge-focused artificial intelligence (AI) workloads. First, we present the main goals of the project, alongside the key enabling technologies for our accelerator, such as a novel augmented silicon photonics platform integrating photonic devices based on (non-volatile) phase-change and III-V materials alongside high-speed photonic components. Then, we discuss the practical use-cases for benchmarking and our currently developed comprehensive simulation framework (based on gem5 and pytorch) for validation.
Featured Image
Photo by Adi Goldstein on Unsplash
Why is it important?
Modern AI is exceedingly capable, but carries significant energy, hardware and carbon footprint costs. It’s now clear that AI can provide significant added value for both the cloud and the edge – but to truly unlock the potential in edge applications, the dedicated hardware running the workloads must be energy efficient without compromising performance. Our photonics-based computing hardware aims to offer both high computational power and increased energy efficiency.
Perspectives
Currently, most of the published papers on optical computing focus mainly on devices and circuits. Meanwhile, system-level studies of photonic AI accelerators are less common. In NEUROPULS, our goal is to deliver and validate a complete photonics-enabled system, including the required peripherals, photonic-electronic interfaces, controllers, and a RISC-V enabled processor. Furthermore, industrial partners taking part in the project help us steer the hardware specifications towards a practically viable AI accelerator architecture.
Matěj Hejda
Hewlett Packard Enterprise Co
Read the Original
This page is a summary of: Invited: Neuromorphic Architectures Based on Augmented Silicon Photonics Platforms, June 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3649329.3665347.
You can read the full text:
Contributors
The following have contributed to this page
