Input-Conditioned Quantisation for ENOB Improvement in CIM ADC Columns Targeting Large-Length Partial Sums

Compute-in-memory (CIM) ADC columns perform Multiply-Accumulate (MAC) operations in the analog domain. Pitch-matching analog-to-digital converters (ADCs) to dense memory pitches limits the area and achievable precision of ADCs. Consequently, CIM literature focuses on increasing MAC effective-number-of-bits (ENOB) by performing quantisation with lower step-size (to reduce error variance), clipping the long tail of the MAC probability density function (PDF) in the process. This brief proposes an alternative, which is to quantise the input-conditioned MAC- PDF (CMPDF), which is the MAC distribution resulting from a given input. A one-shot CMPDF location technique is detailed, which enables the proposed input-conditioned MAC quantisation (ICQ) scheme. Various DNN s are implemented in state-of-the-art clipped quantisation technique (Optimal Clipping Criterion) and compared against ICQ. Results show that DNNs employing ICQ see accuracy improvement over DNNs that do not employ ICQ by 1.6 – 39.4% at iso-ADC precision. DNN layers implemented using ICQ saw an improvement in MAC ENOB up to 4.3 bits over the precision of the employed quantiser (ADC), which is upto 2.2 bits more than the OCC result for the same.