A 2T-1R Cell Array with High Dynamic Range for Mismatch-Robust and Efficient Neurocomputing
Shubham Sahay, Mohammad Bavandpour, Mohammad Reza Mahmoodi, Dmitri B. Strukov
Abstract
Dedicated neuromorphic engines have become indispensable to perform resource-intensive computations involving high-dimensional data on mobile platforms in this era of internet-of-things (IoT) and big data. Vector-by-matrix multiplication (VMM) accelerators form the kernel of these engines. Time-domain based mixed-signal VMM approaches, especially those utilizing emerging non-volatile memory cells as programmable current sources/sinks to realize tunable weights, are arguably the most area- and energy-efficient implementations. However, the cell non-idealities such as capacitive coupling, nonnegligible minimum cell current (I <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</inf> ), limited dynamic range (I <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">max</inf> /I <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</inf> ) of cell currents, drain-induced barrier lowering and susceptibility to mismatch-effects severely degrades their computational precision and limits their area- and energy- efficiency. Therefore, in this work, for the first time, we propose a novel 2T-1R cell that overcomes these limitations and facilitates the effective realization of the true potential of the time-domain VMM implementations. The preliminary estimates indicate that a computational precision of 6-bits and a record energy-efficiency of ~1.5 PetaOp/J with a throughput of ~3.3 TeraOp/s is achievable in a 200×200 time-domain VMM utilizing the proposed 2T-1R cell array, implemented in the 55-nm technology node, including the contribution of the input/output (I/O) and peripheral circuitry.