xnn.gnn.fast.AUTO_POLICY

xnn.gnn.fast.AUTO_POLICY = AutoPolicy(min_edges={'A100': 50000, 'H200': 60000, 'H100': 60000, 'L40S': 30000}, min_edges_float64={'A100': 25000, 'H200': 30000, 'H100': 30000, 'L40S': 0}, default_min_edges=50000, default_min_edges_float64=25000)

Edge counts above which use_fast="auto" uses the kernels. From the MD-step crossover of the kernels’ fixed per-step cost (about 26-38 ms on an A100, 14-21 ms on an H200 and L40S) with the reference’s cost per edge, measured for the production water model, MACE-OFF23 small/medium and MACE-MP-0 medium: the lighter the model, the later the crossover (float32 A100: 19k edges for MP-0 medium, 90k for the light water models), and the thresholds sit between the two. In float64 the reference is slower and the kernels win earlier; on an L40S (weak float64) always.