Large language models rely heavily on linear projections, where multiply-accumulate (MAC) operations dominate both compute and energy consumption. Post-training quantization (PTQ) is a common technique to reduce this burden by mapping weights to lower-precision values, but standard approaches still require multiplications during inference.

The arXiv paper introduces fractional power-of-two quantization, a PTQ method that restricts weights to fractional powers of two. This restriction allows multiplications to be replaced with simple bit shifts, eliminating the need for hardware multipliers entirely. The work targets efficient and effective LLM inference, aiming to cut both computational cost and energy use without requiring retraining.

Because the abstract is brief, the article does not report specific accuracy or speedup numbers. The contribution is the quantization scheme itself, which offers a path toward multiplier-free LLM deployment on resource-constrained hardware.