Analog eDRAM-RRAM compute-in-memory
Published hybrid-memory work for on-device LLM fine-tuning
This IEEE IMW 2025 paper presents analog multilevel eDRAM-RRAM compute-in-memory co-designed with zeroth-order optimization for on-device LLM fine-tuning.
- The paper reports an RRAM-assisted multilevel eDRAM programming scheme and a PVT-robust time-to-digital converter.
- It reports 12× higher bit density for the MLC-eDRAM design than the compared state of the art, with additional density and retention benefits from BEOL In2O3 FETs.