Measured GPU-memory scaling — native / seasonal compute sizing

Generated by scripts/measure_memory_scaling.py. Peak memory of one DINN → integrate → loss → backward pass (torch.compile'd per-step, pre-checkpointing) vs problem size. Memory of a fixed graph is value-independent, so synthetic inputs reproduce a real fit's footprint at matched shapes.

• Device: NVIDIA H200 (139.8 GB total)

• Box: 15-tracer 2-layer, DT=0.25 d, DINN (per-cell 1x1 conv).

• Fit peak = a + b·(cells·steps): a = 13.1 MB, b = 82.9 B per cell·step, R² = 0.99485 (n=20).

Measurements

cells	steps	train peak (GB)	forward-only (GB)	AD overhead x
2,000	50	0.07	0.06	1.1
2,000	100	0.02	0.00	30.7
2,000	200	0.03	0.00	59.4
2,000	400	0.06	0.00	116.9
4,000	50	0.08	0.06	1.3
4,000	100	0.03	0.00	30.7
4,000	200	0.06	0.00	59.5
4,000	400	0.13	0.00	117.1
8,000	50	0.04	0.00	16.3
8,000	100	0.07	0.00	30.7
8,000	200	0.13	0.00	59.5
8,000	400	0.25	0.00	117.1
16,000	50	0.07	0.00	16.3
16,000	100	0.13	0.00	30.7
16,000	200	0.26	0.00	59.5
16,000	400	0.50	0.00	117.1
32,000	50	0.14	0.01	16.3
32,000	100	0.26	0.01	30.8
32,000	200	0.51	0.01	59.6
32,000	400	1.01	0.01	117.2

Extrapolation to the proposal grids

Seasonal assumes ~2000 steps (one annual cycle at DT=0.25 d ≈ 1460 + spin-up); time-mean ~200. Pre-checkpointing peak = the upper bound to size hardware against.

target	cells	steps	predicted peak (GB)
LLC90 global · time-mean	105,300	200	1.6
LLC90 global · seasonal	105,300	2000	16.3
LLC270 global · time-mean	947,700	200	14.6
LLC270 global · seasonal	947,700	2000	146.3
3-AOI native · seasonal (~20k cells)	20,000	2000	3.1