MIVE: A Minimalist Integer Vector Engine for Softmax LayerNorm and RMSNorm Acceleration
Researchers have developed a new hardware architecture called MIVE (Minimalist Integer Vector Engine) designed to accelerate critical operations in large language models (LLMs). MIVE is a programmable engine that can efficiently handle Softmax, LayerNorm, and RMSNorm functions within a single datapath, reducing the need for duplicated hardware resources. An ASIC implementation demonstrated that MIVE offers improved area and hardware efficiency compared to existing standalone accelerators for these operations. AI
IMPACT MIVE's efficient hardware design could lead to faster and more power-efficient inference for large language models.