Publikation
Q-OIG: Quality-Constrained Approximate OIG Synthesis for MAC-Based Logic-in-Memory
Fatemeh Shirinzadeh; Saeideh Nabipour; Abhoy Kole; Kamalika Datta; Saeideh Shirinzadeh; Rolf Drechsler
In: IEEE Nordic Circuits and Systems Conference. IEEE Nordic Circuits and Systems Conference (NorCAS-2026), October 20-21, Tampere, Finland, IEEE, 2026.
Zusammenfassung
Approximate computing improves the efficiency of
error-tolerant systems by trading bounded accuracy for reduced
implementation cost. However, existing approximate logic synthesis techniques primarily target binary graph representations,
leaving native multi-input OR-Inverter Graphs (OIGs) relatively
unexplored for MAC-based logic-in-memory execution.
This paper presents Q-OIG, a quality-constrained approximate
synthesis framework that operates directly on native OIGs. QOIG combines lightweight local transformations with mappingaware archive-based evolutionary search to minimize MAC execution cycles under prescribed quality constraints. Candidate circuits are optimized using stochastic simulation and independently
validated. Experiments on ISCAS-85 and IWLS benchmarks
demonstrate substantial cycle reductions under BER constraints.
An RCA-8 case study further shows competitive cycle–error
trade-offs against a recent MAGIC-based approximate adder in
terms of MAE and MSE.
