A speed comparison can change dramatically when the conventional solver is allowed to use the hardware well. In a benchmark of AC power-flow calculations, researchers found that parallelised physics-based solvers could match or outperform the three neural substitutes they tested while satisfying the electrical equations much more accurately. Throughput among CPU implementations alone varied by hundreds of times, making the chosen baseline central to the result.[1]
Benjamin Donnot and colleagues compared seven CPU-based solvers, two GPU-based solvers and three published machine-learning surrogates across 15 grid models containing 14 to 70,000 buses. A bus is a node connecting elements such as generators and loads. The task was AC power flow: calculate an electrical state consistent with a specified network, generation and demand.[1]
Compare both speed and the electrical residual
The difference between the CPU implementations was roughly 2.5–3 orders of magnitude, or about 316 to 1,000 times. The paper also contrasts a single CPU core with a parallelised implementation. A learned model running on a GPU can look especially fast against one slow serial baseline, while a comparison with a better-engineered alternative tells a different story.[1]
For physical validation, the team sampled 9,000 scenarios per grid—135,000 grid-scenario combinations—and evaluated outputs using an independently implemented Kirchhoff-current-law mismatch check. With five iterations, gpusim2grid’s reported worst-case mismatch was 2.31 × 10⁻¹⁰, below the shared 10⁻⁸ tolerance. The separate throughput experiments measured how many scenarios the methods could process. Keeping both measurements visible prevents an approximate answer from being treated as interchangeable with a tightly converged solution.[1]
The new gpusim2grid solver retains Newton–Raphson numerical methods and batches scenarios on the GPU. That is an engineering route to greater speed without replacing the governing equations with a trained approximation. The public benchmark and evaluator give other developers a common basis for checking those trade-offs on their own systems.[1]
Match the benchmark to the workload
The throughput comparison reused a fixed surrogate architecture across the grid suite without retraining or resizing it for the largest cases; its original training covered grids only up to 2,000 buses. The authors flag that limitation. Their result applies to the tested models and configuration, and a model adapted to a larger grid could perform differently.[1]
Plain power flow is also distinct from choosing the best dispatch or control action. Optimal power flow and grid control add other decisions and constraints, including combinatorial structure, for which learning-based approaches may have different advantages. The benchmark helps locate where a proposed method adds value by keeping the underlying task precise.[1]
For an operator assessing a speed claim, the comparison needs four visible ingredients: the exact power-system problem, the conventional solver, the computing resources used by both methods and a common check on their answers. Batching and data-transfer costs also need to match the intended workload. The best tool for a large offline scenario sweep may differ from the best tool for one time-critical calculation.[1]
The benchmark’s contribution is a stronger reference point. A new learned solver can now be compared with fast CPU and GPU alternatives under the same electrical check. That makes an eventual improvement more useful: it has to survive both competent numerical competition and the physics of the grid.[1]
Sources & further reading
- Benchmarking CPU, GPU, and ML Power-Flow Solvers from 14 to 70,000 Buses RTE researchers / arXiv — author manuscript copy · 20 Oct 2026 · Accessed 25 Oct 2026