Software toolchain for the hybrid digital-analog, memristor-based accelerator for machine learning

The deceleration of transistor feature size scaling has motivated the growing adoption of specialized accelerators implemented as GPUs, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), and more recently new types of computing such as neuromorphic, ultra-low energy, reversible, stochastic, optical, quantum, combinations, etc. There’s a tension between specialization and generalization, with the current state trending toward special purpose, optimized accelerators.

Dejan Milojicic examines the software toolchain for a hybrid digital-analog, memristor-based accelerator, the Dot Product Engine (DPE). He explores the toolchain for a select pattern of applications in machine learning and demonstrates that the software toolchain makes an accelerator, such as the DPE, more general, which enables broader adoption and better utilization. More specifically, you’ll learn about the compiler, assembler, loader, application optimizer, ONNX plugins, and device driver for the DPE.