Towards Memory-Efficient Processing-in-Memory Architecture for Convolutional Neural Networks

Convolutional neural networks (CNNs) are widely adopted in artificial intelligent systems. In contrast to conventional computing-centric applications, the computational and memory resources of CNN applications are mixed together in the network weights. This incurs a significant amount of data movement, especially for high-dimensional convolutions. Although recent embedded 3D-stacked Processing-in-Memory (PIM) architecture alleviates this memory bottleneck to provide fast near-data processing, memory is still a limiting factor of the entire system. An unsolved key challenge is how to efficiently allocate convolutions to 3D-stacked PIM to combine the advantages of both neural and computational processing. This paper presents Memolution, a compiler-based memory-efficient data allocation strategy for convolutional neural networks on PIM architecture. Memolution offers thread-level parallelism that can fully exploit the computational power of PIM architecture. The objective is to capture the characteristics of neural network applications and present a hardware-independent design to transparently allocate CNN applications onto the underlining hardware resources provided by PIM. We demonstrate the viability of the proposed technique using a variety of realistic convolutional neural network applications. Our extensive evaluations show that, Memolution significantly improves performance and the cache utilization compared to the baseline scheme.