High-Performance Data Mapping for BNNs on PCM-based Integrated Photonics

TL;DR

This work targets the latency and data-movement bottlenecks of state-of-the-art DNN hardware by focusing on Binary Neural Networks (BNNs) implemented with Computation-In-Memory (CIM) and photonic memories. It introduces TacitMap, a highly parallel data mapping that enables 1-step XNOR+Popcount on CIM crossbars, and EinsteinBarrier, an oPCM-based accelerator that leverages wavelength division multiplexing (WDM) to further boost parallelism. The combination yields substantial latency reductions (up to $154\times$ for TacitMap and up to $3113\times$ for EinsteinBarrier) with energy within 60% of the CIM baseline, demonstrated across six BNNs on MNIST and CIFAR-10. This work demonstrates the viability of integrating oPCM photonics with CIM to realize ultra-fast, energy-efficient BNN hardware, and points to future explorations of multi-bit PCM and broader hardware-software mappings.

Abstract

State-of-the-Art (SotA) hardware implementations of Deep Neural Networks (DNNs) incur high latencies and costs. Binary Neural Networks (BNNs) are potential alternative solutions to realize faster implementations without losing accuracy. In this paper, we first present a new data mapping, called TacitMap, suited for BNNs implemented based on a Computation-In-Memory (CIM) architecture. TacitMap maximizes the use of available parallelism, while CIM architecture eliminates the data movement overhead. We then propose a hardware accelerator based on optical phase change memory (oPCM) called EinsteinBarrier. Ein-steinBarrier incorporates TacitMap and adds an extra dimension for parallelism through wavelength division multiplexing, leading to extra latency reduction. The simulation results show that, compared to the SotA CIM baseline, TacitMap and EinsteinBarrier significantly improve execution time by up to ~154x and ~3113x, respectively, while also maintaining the energy consumption within 60% of that in the CIM baseline.

Figures (8)

• Figure 1: CIM support of VMMs in NNs.
• Figure 2: Concepts of TacitMap vs CustBinaryMap hirtzlin2020digital-BNN_differential_SA-MahdiBCIM24.
• Figure 3: TacitMap vs CustBinaryMap data mapping.
• Figure 4: EinsteinBarrier system placement and overview.
• Figure 5: WDM in oPCM core.