MIVE: A Minimalist Integer Vector Engine for Softmax LayerNorm and RMSNorm Acceleration

The article introduces the Minimalist Integer Vector Engine (MIVE), a programmable architecture designed to accelerate non-linear vector normalization operations such as LayerNorm, RMSNorm, and Softmax within a unified datapath. By leveraging common computational patterns, MIVE enhances hardware sharing and reduces implementation overhead, resulting in higher area and hardware efficiency compared to existing standalone accelerators. This development is significant for practitioners as it addresses critical bottlenecks in LLM inference, offering a more efficient solution for specialized hardware accelerators.