Researchers baked 3x inference speedups directly into LLM weights — without speculative decoding

Researchers from the University of Maryland, Lawrence Livermore, Columbia and TogetherAI have developed a training technique that triples LLM inference speed without auxiliary models or infrastructure ...
IEEE

High-Performance Matrix Multiplication Using Block Parallelization on CPU and GPU

Abstract: This paper examines the performance issues associated with computing devices performing arithmetic operations on large matrices. One of the optimal methods for matrix multiplication is to ...