Breaking01 / AI
GPU-CPU method improves optimisation scores by up to 4.9%
The method passes GPU-calculated variable values to CPU solvers without training data or a pretrained model.

Researchers reported that a GPU-CPU hybrid method improved average solution scores by up to 4.9% in time-limited optimisation tests, according to their paper's abstract published on 13 Nov. The gain is a researcher-published benchmark figure.
The abstract identifies the task as Weighted Partial MaxSAT (WPMAX-SAT), an optimisation problem that combines mandatory constraints with optional conditions carrying different weights.
The researchers said the GPU adjusted variables for each problem without a training dataset or a pretrained predictive model. They designed the loss function, which directs the optimisation, to prioritise satisfying mandatory constraints before improving the weighted score for optional conditions.
The abstract describes two ways to pass GPU results to CPU solvers: adding selected variable values as extra conditions, or using the full set of values as a starting point for the search. The latter guides the CPU solver's search while preserving the original problem.
The researchers evaluated 229 weighted problems from the MSE 2024 benchmark with a 300-second time limit. Their published figures showed average solution-score improvements of up to 4.9% over tt-open-wbo-inc-Glucose and up to 4.8% over SPB-MaxSAT-c-FPS.
The researchers did not disclose the GPU model or its power consumption in the abstract.