A 709.3 TOPS/W Event-Driven Smart Vision SoC with High-Linearity and Reconfigurable MRAM PIM
Wenao Xie, Haoyang Sang, Beomseok Kwon, Dongseok Im, Sangjin Kim, Sangyeob Kim, Hoi‐Jun Yoo
Abstract
This paper presents a high energy-efficiency system-on-chip (SoC) for smart vision applications. A novel reconfigurable and high-linearity MRAM PIM is proposed and achieve high-accuracy event detection (ED) and high-energy-efficiency inference in the event-driven system. The 28nm SoC obtains an average core power consumption of 7.3mW, with an average power reduction of 30.9% at all event densities. The MRAM PIM achieves the peak energy efficiency of 709.3 TOPS/W (1bI-1bW-3bO) at 200MHz and inference accuracy of 91.90%/67.61% (CIFAR-10/100) with ResNet-20.
Topics & Concepts
Magnetoresistive random-access memoryLinearityComputer scienceEfficient energy useEnergy consumptionEvent (particle physics)Energy (signal processing)Power consumptionReduction (mathematics)InferenceLookup tableChipPower (physics)Embedded systemComputer hardwareElectronic engineeringArtificial intelligenceEngineeringMathematicsRandom access memoryElectrical engineeringPhysicsStatisticsQuantum mechanicsGeometryTelecommunicationsProgramming languageCCD and CMOS Imaging SensorsAdvanced Memory and Neural ComputingNeuroscience and Neural Engineering