Session

DescriptionIn this session, we explore innovative strategies that combine alternate arithmetic approaches, novel architectures, and co-design techniques to enhance AI systems. Focusing on both theoretical and practical advancements, the session highlights how new arithmetic methods and specialized hardware, including sparsity handling and approximation, are driving improvements in performance and efficiency for AI models. More specifically, the session presents various microarchitectures for efficient processing such as bit-sparse attention architecture, bit-level sparse MAC architecture, and dense-sparse systolic array architecture, and novel arithmetic such as Polar form and floating-point approximation. It also includes emerging accelerator architectures including Gaussian Splatting and Point-Cloud Processing.