The Potential of Efficient Photonic Multiply-Accumulator Processors

Novel photonic compute concepts promise to outperform conventional electronic hardware in terms of throughput and energy efficiency for multiply-accumulate operations (MACs). Several on-chip solutions have been demonstrated such as photonic crossbar arrays [1], Mach-Zehnder interferometers [2], or microring resonators [3]. However, it is difficult to compare their performances among each other, due to inconsistent figures of merit, e.g. for the systems' energy consumption. In this work, we review existing photonic and conventional compute units, introduce a systematic analysis of the overall energy consumption of photonic compute modules, and compare their performance regarding energy efficiency and throughput. It can be seen that digital-analog as well as electro optic conversions dominate energy needs in analog electro-optic compute architectures. Nevertheless, scaling up the kernel size of optical computers will reduce the impact of conversions and lead to an energy advantage over conventional compute units, making them promising candidates for artificial intelligence applications.