15.4 A 22nm 2Mb ReRAM Compute-in-Memory Macro with 121-28TOPS/W for Multibit MAC Computing for Tiny AI Edge Devices

Abstract

Nonvolatile computing-in-memory (nvCIM) can improve the latency (t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AC</sub> ) and energy-efficiency (EF <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MAC</sub> ) of tiny AI edge devices performing multiply-and-accumulate (MAC) computing after system wake-up. Prior nvCIMs have proven effective for binary input (IN) and weight (W), and 3b output (OUT) [1], 1-8-1b IN-W-OUT [2], and 2-3-4b IN-W-OUT [3] neural networks; however, the higher precision (4-4b IN-W) for MAC operations is needed for multi-bit CNNs to achieved high-inference accuracy [4]. As Fig.15.4.1 shows, improving the precision of nvCIM macros involves various challenges. (1) A large number of activated WLs provides a wide range of BL current (I <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BL</sub> ) resulting in an inaccurate BL-clamping voltage (V <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BLC</sub> ); as well as a large (I <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BL</sub> ) requiring a large array area due to the need for wide metal lines to support high-current density. (2) Previous “WL = input” approaches suffer from: (a) few parallel inputs (IN#) due to (1), and (b) long (t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AC</sub> ) in multiple cycles of binary WL inputs on 1T1R cells for multibit inputs. (3) Previous positive-negative-split weight-mapping consumes high total (l <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BL</sub> ) and area overhead (needing 2x(m-1) cells for a signed m-bit weight) for cell arrays with high-weight precision. (4) Long (t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AC</sub> ) and a large number of reference currents (IREF#) for high-precision outputs. To overcome these challenges, this work proposes: (1) a BL-IN-OUT multibit computing (BLIOMC) scheme using a single WL-on and input-aware multibit BL clamping (IA-MBC) to shorten (l <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BL</sub> ) for multibit inputs, increase IN#, and reduce the (l <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BL</sub> ) range/size for accurate (V <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BLC</sub> ) and a compact array area. (2) Scrambled 2's complement (S2C) weight mapping (S2CWM), input-aware source-line (SL) voltage biasing (IA-SLVB), and an S2C value combiner (S2CVC) to reduce area overhead and l <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">BL</sub> in the cell array. (3) A dual-bit small-offset current-mode sense amplifier (DbSO-CSA) to reduce IREF# and t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AC</sub> . A fabricated 22nm 2Mb ReRAM-CIM macro presents the first 4b-input nvCIM macro, featuring a 9.8-18.3ns t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">AC</sub> and an EF <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MAC</sub> of 121.3-28.9TOPS/W from binary to 4bIN-4bW-11bOUT compute precisions.