Measured wall-clock scaling — native / seasonal compute timing¶
Generated by scripts/measure_compute_time.py. Median wall-clock of one DINN → integrate → loss → backward epoch (eager autograd) vs problem size. Timing of a fixed graph is value-independent, so synthetic inputs reproduce a real fit's per-epoch cost at matched shapes.
-
Device: NVIDIA GeForce RTX 5090 Laptop GPU (23.9 GB)
-
Per-epoch cell-fit
t = a + b·cells@ 200 steps: a = 1827.6 ms, b = 576 ns/cell, R² = 0.4581 (n=8).
Measurements (one optimizer epoch = fwd + bwd)¶
| cells | steps | train (median) | per-step | per-cell·step | fwd-only |
|---|---|---|---|---|---|
| 10,000 | 50 | 461.2 ms | 9.225 ms | 922.47 ns | 195.4 ms |
| 10,000 | 100 | 943.2 ms | 9.432 ms | 943.20 ns | 395.4 ms |
| 2,853 | 200 | 1.82 s | 9.077 ms | 3181.61 ns | 764.1 ms |
| 5,000 | 200 | 1.81 s | 9.072 ms | 1814.31 ns | 763.1 ms |
| 10,000 | 200 | 1.83 s | 9.171 ms | 917.10 ns | 779.5 ms |
| 10,000 | 200 | 1.89 s | 9.428 ms | 942.81 ns | 777.8 ms |
| 20,000 | 200 | 1.83 s | 9.146 ms | 457.28 ns | 783.0 ms |
| 38,809 | 200 | 1.84 s | 9.203 ms | 237.14 ns | 769.6 ms |
| 70,000 | 200 | 1.85 s | 9.262 ms | 132.32 ns | 785.5 ms |
| 105,300 | 200 | 1.90 s | 9.500 ms | 90.22 ns | 774.8 ms |
| 10,000 | 400 | 3.75 s | 9.363 ms | 936.29 ns | 1.56 s |
| 10,000 | 800 | 7.63 s | 9.534 ms | 953.37 ns | 3.11 s |
| 10,000 | 1600 | 14.75 s | 9.216 ms | 921.62 ns | 6.17 s |