| PhD Viva

Name of the Speaker: Mr. Ashwin Balagopal (EE17D200)
Guide: Dr. Janakiraman Viraraghavan
Online meeting link: https://meet.google.com/uof-xuov-guz
Date/Time: 11th April 2025 (Friday), 10.15 AM
Title: Input Conditioned Quantisation For Compute-In-Memory

Abstract :

Artificial intelligence (AI) workloads are increasingly run on energy-constrained edge devices like smartphones, where the von-Neumann (VN) architecture faces performance limitations due to the memory bottleneck—where memory access dominates latency and energy consumption. Compute-in-memory (CIM) offers an energy-efficient alternative by executing Multiply-Accumulate (MAC) operations in the analog domain, but its effectiveness is constrained by the limited area and precision of analog-to-digital converters (ADCs). Traditional approaches like Optimal Clipping Criterion (OCC) attempt to improve MAC precision by clipping the MAC probability density function (PDF), but they overlook the input-specific nature of these distributions. This talk summarises input-conditioned quantisation (ICQ), a method that tracks and tailors references to the input-conditioned MAC PDF, improving the fidelity of quantisation, measured by effective-number-of-bits (ENOB). Software experiments show ICQ improves MAC ENOB by up to 4.3 bits and DNN accuracy by 1.6–39.4% over OCC, with ≤1% accuracy loss on MNIST, CIFAR10, and CIFAR100. Hardware measured results realise MAC ENOB improvement by 1-2.5 bits over the ENOB of the 7-bit on-chip ADC, and inference results on MNIST, CIFAR10, and CIFAR100 are shown with ≤ 1% accuracy loss from the software baseline.