A Chiplet-Based Generative Inference Architecture with Block Floating Point Datatypes

Explore a comprehensive conference talk on chiplet-based generative inference architecture and block floating point datatypes for AI acceleration. Delve into modular, spatial CGRA-like architectures optimized for generative inference, and learn about deep RL-based mappers in compilers for spatial and temporal architectures. Discover weight and activation quantization techniques in block floating point formats, building upon GPTQ and SmoothQuant, and their implementation in PyTorch. Examine an extension to EL-attention for reducing KV cache size and bandwidth. Gain insights from speaker Sudeep Bhoja in this SPCL_Bcast #38 recording from ETH Zurich's Scalable Parallel Computing Lab, featuring an in-depth presentation followed by announcements and a Q&A session.