Joint Hardware-Workload Co-Optimization for In-Memory Computing Accelerators

Mar 4, 2026 by O. Krestinskaya, M. E. Fouda, A. Eltawil, Khaled N. Salama

We built a joint hardware-workload co-optimization flow that rigs evolutionary search to design generalized in-memory computing accelerators that actually work well across many networks instead of just one, and it slashes EDAP by up to ~76–95% across RRAM and SRAM IMC targets. If you care about deployable IMC platforms that trade off workloads explicitly rather than overfitting to a single model, this framework (code included) shows the practical gains and robustness you'd expect but rarely see.

source S2