Performance Benchmarking: Software for the Density Matrix Renormalization Group

The Density Matrix Renormalization Group (DMRG) algorithm is a cornerstone of quantum physics research, used to study complex quantum systems. Despite its importance, the landscape of DMRG software is fragmented, with over 50 distinct implementations. Choosing the right tool is difficult because existing performance evaluations are typically isolated, lack standardized metrics, and often prioritize theoretical features over practical computational efficiency. This paper addresses this gap by proposing a rigorous, performance-oriented benchmarking framework.

The authors argue that meaningful performance evaluation must move beyond simple metrics like "sweep time." Instead, they propose a framework that jointly evaluates computational cost and solution quality. Their approach emphasizes five requirements: joint evaluation of cost and accuracy, use of end-to-end performance metrics, consistent model problems, controlled parameter configurations, and standardized hardware/software environments. By focusing on the interplay between these variables, the framework allows researchers to understand the true trade-offs between different software packages.

Applying this framework to eight different DMRG implementations, the authors reveal significant performance disparities. They find that performance differences can reach up to two orders of magnitude. Crucially, these differences arise not only from the choice of software implementation but also from how users configure parameters—such as bond dimension, eigensolver settings, and truncation criteria. The study demonstrates that systematic evaluation is essential for identifying non-obvious trade-offs, such as the impact of mixed-precision strategies or symmetry-adapted optimizations on overall time-to-solution.