Hybrid electronic-photonic architectures for brain-inspired computing

As artificial intelligence (AI) proliferates, hardware systems that can perform inference at ultralow latency, high precision and low power are crucial and urgently required to deal – especially quasi-locally, i.e. ‘in the edge’ – with massive and heterogenous data, respond in real time and avoid unintended consequences and function in complex and often unpredictable environments. Conventional digital electronics and the associated computer architecture is unable to meet these stringent requirements with sub-ms latency inference and a sub-10W power budget, using convolution neural networks (CNNs) on benchmarks such as ImageNet classification. HYBRAIN’s vision is to realize a pathway for a radical new technology with ultrafast (~1 microsecond) and energy-efficient (~1 watt) edge AI inference based on a world-first, brain-inspired hybrid architecture of integrated photonics and unconventional electronics. The deeply entwined memory and processing like in the mammalian brain obviates the need to shuttle around synaptic weights. The most stringent latency bottleneck in CNNs is in the initial convolution layers. Our approach will take advantage of the ultrahigh throughput and low latency of photonic convolutional processors (PCPs) employing novel phase-change materials in these initial layers to radically speed up processing. Their output is processed using cascaded electronic linear and nonlinear classifier layers, based on memristive (phase-change memory) crossbar arrays and dopant network processing units, respectively. HYBRAIN’s science-towards-technology breakthrough brings together the world’s top research groups from academia and industry in complementary technology platforms. Each of these platforms is already highly promising, but by integrating them, HYBRAIN will have a transformative effect of overcoming existing barriers of latency and energy consumption and will enable a whole new spectrum of edge AI applications throughout society.