Towards Efficient Hyperdimensional Computing Using Photonics

Avi Mohan, Ajay Joshi, Mohsen Imani, Darius Bunandar

Abstract:

Over the past few years, photonics-based computing has emerged as a promising alternative to CMOS-based computing for Machine Learning (ML)-based applications, in particular, for Deep Neural Networks (DNNs). Unfortunately, the non-linear operations and the high-precision requirements of DNNs make it challenging to design efficient photonics-based computing systems for DNNs. Hyperdimensional Computing (HDC) is an emerging, brain-inspired ML technique that enjoys several advantages over DNNs, including (i) being lightweight, i.e., involving simple linear algebraic operations; (ii) requiring low-precision operands; and (iii) being robust to noise introduced by the nonidealities in the hardware.
In this paper, we argue that photonic computing and HDC complement each other effectively. We propose PhotoHDC, the f irst electro-photonic accelerator for HDC training and inference, supporting traditional, record-based, and graph encoding schemes. Our novel accelerator microarchitecture utilizes MachZehnder Modulators and photodiodes to accelerate HDC encoding, bundling, and similarity measurement operations while minimizing the use of digital modules for other system components. Our analyses show that PhotoHDC can achieve two to five orders of magnitude lower Energy-Delay Product (EDP) than the stateof-the-art electro-photonic DNN accelerators when running HDC training and inference across popular datasets. We also show that photonics can overcome the challenges faced by Computein-Memory-based HDC accelerators and achieve multiple orders of magnitude lower EDP for both HDC training and inference.