Preprint reports up to 32% decode energy savings from GPU clock control

A preprint reports that locking a GPU's processing clock recovered up to 32% of the energy used to generate language-model output, with minimal throughput loss. Researchers tested four attention architectures on an NVIDIA H200 using vLLM eager execution. During output generation, the tested workloads waited on high-bandwidth memory and idled between kernel launches, drawing less than half the GPU's 700W rating. Ordinary power caps therefore had little effect on that phase. The cap engaged only when a large batch ran a long generation, limiting how broadly the finding applies across serving workloads.