Performance Enhancement of Edge-AI-Inference Using Commodity MRAM: IoT Case Study

In this paper we demonstrate how performance of Edge-AI-Inference hardware can be enhanced by effectively using emerging commodity MRAM chips inside specific accelerator pipelines. We consider the special case of IoT-centric ‘normally-off/low-frequency’ AI inference workloads to benchmark the proposed approach. The proposed Non-Volatile AI Inference Accelerator (NVIA) is realized using FPGA and off-the-shelf MRAM chips. NVIA is benchmarked for Human activity recognition (HAR) dataset. Significant power gains, ∼9x (with Toggle-MRAM (180 nm)) and ∼750x (with STT-MRAM (22 nm)), were achieved compared to volatile SRAM.