Open-RIMC: Open-Source RRAM-based IMC Co-Processor with Reconfigurable Logic Mapping

In this study we present Open-RIMC, an open-source RRAM-based IMC (In-memory computing) co-processor that enables energy-efficient mapping of MAC (multiply and accumulate) operations while providing reconfigurable logic-mapping. The proposed RRAM-IMC IP is implemented using the open-source Skywater-130 nm PDK and a fully open-source EDA tool-chain. Our IMC block acts as a co-processor to the RISC-V based Caravel co-processor. We demonstrate BNN (Binary Neural Network) as a potential application along with analysis on impact of variability on inference performance of the IMC-macro. We further analyze trade-offs due to logic-mapping. The design is open-sourced and available online at https://gitlab.com/NVM_IITD_Research/RRAM_IMC_IITD. The proposed design can be used for realizing efficient edge AI (artificial intelligence) quantized neural networks (QNN).