Litcius/Paper detail

All-Digital Time-Domain Compute-in-Memory Engine for Binary Neural Networks With 1.05 POPS/W Energy Efficiency

Jie Lou, Christian Lanius, Florian Freye, Tim Stadtmann, Tobias Gemmeke

2022ESSCIRC 2022- IEEE 48th European Solid State Circuits Conference (ESSCIRC)19 citationsDOI

Abstract

This paper presents an all-digital time-domain compute-in-memory (TDCIM) engine for binary neural networks (BNNs), which is based on commercial standard cells facilitating technology mapping. The proposed TDCIM engine exploits energy-efficient computing principles, supports data reuse and employs double-edge triggered operation. Time domain wave-pipelining technique is also introduced to improve throughput by 1.5x while preserving accuracy. We use Structured Data-Path (SDP) placement and custom routing flow during place and route (P&R) to reduce systematic variations. The measured arrival time of different MAC results is sufficiently bounded to preserve accuracy across PVT variations. Fabricated in a 22nm process, the proposed BNN engine can achieve an energy efficiency of 1.05 POPS/W at 0.5V matching the accuracy of the PyTorch baseline of 99.14% on the MNIST dataset.

Topics & Concepts

MNIST databaseComputer scienceDomain (mathematical analysis)Binary numberThroughputEfficient energy useBenchmark (surveying)Matching (statistics)Process (computing)ReuseEnergy (signal processing)Routing (electronic design automation)Bounded functionArtificial neural networkParallel computingComputer engineeringReal-time computingEmbedded systemArtificial intelligenceMathematicsEngineeringElectrical engineeringTelecommunicationsGeographyArithmeticStatisticsGeodesyMathematical analysisWaste managementWirelessOperating systemAdvanced Memory and Neural ComputingFerroelectric and Negative Capacitance DevicesSemiconductor materials and devices