MEMA Runtime Framework: Minimizing External Memory Accesses for TinyML on Microcontrollers

Learn about the MEMA framework for optimizing TinyML systems in this 13-minute research symposium presentation from Harvard University PhD students. Explore how this innovative framework analytically determines optimized schedules and kernels to minimize external memory access during matrix multiplication operations on microcontrollers. Discover the framework's ability to account for hardware constraints and problem sizes while avoiding time-consuming heuristic searches. Examine performance comparisons with existing state-of-the-art libraries, including impressive results showing up to 1.8x speedup and 44% energy reduction compared to CMSIS-NN when running neural network benchmarks on the ARM Cortex-M4.