SISA: A Scale-In Systolic Array for GEMM Acceleration
The episode discusses SISA, an architecture for General Matrix-Matrix Multiplication (GEMM), which is core to LLMs. It addresses the inefficiency of traditional accelerators when handling skewed or small input data. SISA uses a scale-in approach to achieve up to an eight point five two times speedup and a ninety-three percent reduction in energy consumption.