BWA-NIMC: Budget-based Workload Allocation for Hybrid Near/In-Memory-Computing
DescriptionTo enable efficient computation for convolutional neural networks (CNNs), near-memory-computing (NMC) and in-memory-computing (IMC) are proposed to improve energy efficiency and throughput. However, CIM is influenced by the process variation of non-volatile memory (NVM) and nonlinearity in the I-V curve, while it exhibits higher efficiency than NMC. In this work, we propose BWA-NIMC to incorporate the advantages of NMC and IMC. We exploit the inherent bit-level importance and apply hardware-aware selection metrics to systematically allocate workloads between NMC and IMC within arbitrarily targeted resource budgets. Simulation results show that BWA-NIMC improves the accuracy by 18.38-45.18% under limited constraints.
TimeWednesday, July 12th4:10pm - 4:25pm PDT
Location3003, 3rd Floor
In-memory and Near-memory Computing Architectures, Applications and Systems