Systematic GPU Benchmarking of CRM

Conduct a systematic benchmark of Centroid-Referenced Mahalanobis Matching and its computational components on GPU-accelerated backends such as CuPy or JAX to assess practical performance at large scale.

Background

Centroid-Referenced Mahalanobis Matching relies primarily on dense linear-algebra operations, including covariance estimation, Cholesky factorization, and vectorized matrix–vector products. The paper notes that these operations are compatible with GPU-accelerated implementations, but the reported computational evaluations use a single CPU core. A systematic GPU benchmark is therefore left unresolved and would clarify whether GPU execution materially improves CRM's scalability and runtime relative to CPU implementations and competing matching procedures.