SimAMC tackles the problem of simulating closed-loop analog matrix computing circuits with non-idealities, solving matrix inversion and eigenvector problems in the presence of device programming error, data conversion error, thermal noise, op-amp input offset, and interconnect resistance. This is the first simulator I've seen that handles closed-loop AMC circuits at this level of fidelity while staying fast enough for iterative design work.
Why Closed-Loop AMC Simulation Is Harder
Open-loop analog matrix-vector multiplication has well-understood models. Closed-loop circuits that solve matrix equations like $Ax = b$ or compute eigenvectors are a different beast. They are inherently more complex and far more sensitive to every non-ideality in the analog chain. A small programming error in a resistive memory cell can throw off the entire solution. Simulating that with SPICE on realistic matrix sizes takes forever.
The SimAMC team built an alternating iterative algorithm specifically for real-valued matrix computing circuits. Instead of brute-force transient simulation, they iterate between the analog domain model and a digital correction loop, capturing the dominant non-idealities without simulating every transistor.
Validation Against SPICE and the Speedup
SimAMC's results match SPICE with excellent agreement on matrix inversion and eigenvector solving benchmarks. The paper reports a speedup of several orders of magnitude over SPICE. That translates from days to minutes for a 64x64 matrix, making it practical to sweep design parameters and corner cases that were previously too expensive to simulate.
They validated across all five listed non-idealities: device programming error, data conversion quantization and noise, thermal noise, operational amplifier input offset voltage, and interconnect resistance. Each non-ideality can be toggled independently, letting designers isolate which imperfection dominates accuracy loss in their circuit.
What This Enables Next
With SimAMC, engineers can now run monte carlo simulations of closed-loop AMC accelerators before taping out. Early-stage design optimization for resistive memory arrays just got a lot more practical. Expect follow-up work that extends the simulator to complex-valued matrices or extends the non-ideality library to include device aging and temperature drift.
Source: SimAMC: A Fast and Accurate Simulator for Resistive Memory-Based Analog Matrix Computing with Non-Idealities
Domain: arxiv.org
Comments load interactively on the live page.