FLAMES: Framework for Learning With Analog MEmory Systems

Online training is central for increasing the speed and accuracy of edge ML. In this paper, a novel framework is presented to execute training and inference operations on-chip using analog memory. With the proposed system, no data communication between memory and the processor is needed. In addition, the framework is designed in fully analog domain, eliminating the need for using expensive analog-to-digital and digital-to-analog converters. Finally, the same circuit is utilized for both pairwise training and inference. To evaluate the performance within the computational constraints of Cadence, a reduced ten-class MNIST dataset is generated by down-sampling the MNIST digit images from 784 features to 16 features. A total of 45 systems are trained to perform the individual pairwise classifications of MNIST data. The proposed system exhibits an average accuracy of 91.14% exceeding the performance of a digital 6-bit resolution system. The proposed framework is designed and demonstrated in Cadence Virtuoso with TSMC 65 nm technology node. The pairwise training and inference circuit occupies an area of 0.243 mm 2 and consumes 6.593 fJ energy per each multiply-accumulate (MAC) operation.