Session

DescriptionThe success of transformer models has been demonstrated across various ML-based services in the fields of natural language processing, computer vision, video processing, etc. Such versatile applications also introduce challenges stemming from their vast demand for computation, memory capacity, and memory bandwidth, alongside concerns regarding their implications for energy consumption and potential model misuse. This session presents technical solutions and scientific advancements through behavioral analysis of input tokens, novel quantization algorithms, hardware architecture techniques, and hardware-software co-design. The discussions aim to tackle the pressing issues faced by transformer models, making their operations more practical and efficient for real-world applications.