Exploring Algorithm Frameworks for Compute-in-Memory Accelerators
Compute-in-memory accelerator design requires selecting training configurations before circuit- and system-level implementation, while the relevant device nonidealities are contract-specific. We present a hardware-calibrated noise-contract-driven design-space-exploration methodology that screens training viability across noise intensity, noise composition, architecture, and dataset axes. The contract separates total forward perturbation from four weight-channel terms and uses a one-sided robustness ratio only as an empirical screening indicator. Across nine controlled strategy vehicles, the loop returns viability classes, failure boundaries, implementation-dependence evidence, and protocol-bounded design rules. The screening identifies a staged dynamic schedule with no measurable benefit within its tested six-seed budget, no gain from the contrastive variant on Tiny-ImageNet-200, descriptive label-noise differences of at most 0.15 percentage points across twelve single-seed combinations, and a contract- and implementation-dependent VGG16 failure boundary. All hardware-level evidence is simulator-based and calibrated to vendor-reported characteristics; no physical-chip measurement, silicon correlation, or PPA Pareto frontier is claimed. The method is intended as a pre-implementation screening layer that complements circuit- and system-level CIM evaluation.
Authors
- Jian Xia
- Wsifeng Jin
- Yishuai Duan
- Fan Liu
- Rongtao Fu
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24422459
- Primary Topic
- Advanced Memory and Neural Computing
- Type
- preprint