A Broad-Spectrum and High-Throughput Compression Engine for Neural Network Processors

Feature map (fmap) compression is an effective approach for alleviating the memory access bottleneck. However, prior works are limited to specific networks and data types, and they also suffer from limited throughput. In this brief, we propose a lossless algorithm-architecture co-optimized compression engine for fmaps. At the algorithm level, we design a broad-spectrum adaptive compression algorithm. It dynamically selects compression parameters and is suitable for various kinds of networks and data types. Based on the algorithm, a multi-lane compression architecture with inter-lane decoupling and workload balancing is designed to improve throughput and area efficiency. Compared to the State of the Art work, our design achieves significant compression ratio improvements across nearly all the benchmarks, while also reducing area cost by 17% and increasing peak throughput by
$3\times $
.