Stream-Based Computing for Mitigating RRAM Nonlinearity in Compute-in-Memory 8 views

Resistive random-access memory (RRAM) crossbars have emerged as a promising platform for compute-in-memory (CiM) architectures by performing analog matrix–vector multiplication (MVM) directly within memory, reducing the data movement that limits the performance and energy efficiency of conventional computing systems. However, practical RRAM devices exhibit nonlinear current–voltage (I–V) characteristics. When inputs are encoded as analog voltage amplitudes, this nonlinearity introduces input-dependent computational error that degrades MVM accuracy. This thesis investigates Asynchronous Stream Computing (ASC), a stream-based computing paradigm, as a circuit-level approach for mitigating RRAM nonlinearity in analog CiM. As a prerequisite, a comprehensive survey of eight widely used RRAM compact models is conducted to identify a model that accurately captures read-regime nonlinear behavior while remaining compatible with circuit-level simulation. The Stanford Verilog-A model is selected, calibrated, and validated in Cadence Virtuoso through reproduction of switching hysteresis, verification of a non-destructive read operating region, and characterization of a stable conductance window. Using the validated device model, an ASC-based CiM architecture is evaluated in which input values are represented by the duty cycle of constant-amplitude asynchronous ΣΔ streams. Because each RRAM cell is always read at a fixed bias voltage, the nonlinear I–V characteristic has no direct impact on the computed output. Circuit-level simulations demonstrate that the proposed approach reduces the maximum single-cell relative error from approximately 25% to below 4%, achieving sub-1% error in the high-resistance state. For 4×1, 8×1, and 16×1 crossbar arrays, the mean normalized MVM error is reduced by up to 387× compared with conventional voltage-amplitude encoding, while providing a configurable accuracy–latency tradeoff through the accumulation window. As an exploratory extension, this thesis also investigates stream-based operation of closed-loop RRAM crossbars for in-memory solution of linear systems, demonstrating the relationship between inversion accuracy and matrix conditioning. Overall, these results demonstrate that asynchronous stream encoding provides an effective circuit-level mechanism for mitigating RRAM nonlinearity, improving analog MVM accuracy without requiring device modifications or nonlinearity-aware training.

MS (Master of Science)

Resistive RAM (RRAM), Compute-in-Memory (CiM), Matrix–Vector Multiplication (MVM), Asynchronous Stream Computing (ASC), Σ∆ Modulation

English

All rights reserved by the author (no additional license for public reuse)

2026-07-17