Benchmark note · 17 September 2026
How the Studio scales across the CMMA.
CMMA means Costa de Manguezais de Macromaré da Amazônia — the Amazon macro-tidal mangrove coast. This note compares continuous and persistent-block processing with a complete land-cover grid and ANADEM, a 30 m Digital Terrain Model (MDT/DEM) that represents terrain height.
The benchmark covers one annual transition from 2024 to 2025 across the complete CMMA grid. The land-cover input describes classes such as mangrove, water, natural vegetation, and blocked areas. ANADEM supplies the matching terrain height: it is a Digital Terrain Model (MDT/DEM), not a land-cover map. Each engine was run once with the same inputs and parameters; annual GeoTIFF snapshots were disabled so the comparison measures simulation work rather than export time.
| Engine | Configuration | Elapsed time | Throughput | Peak RSS |
|---|---|---|---|---|
| Continuous | Active grid held in memory | 123.6 s | 0.631M cell-updates/s | 5.92 GiB |
| Blocks | Persistent 10,000-cell chunks | 476.7 s | 0.164M cell-updates/s | 10.01 GiB |
Input and scenario
CMMA V2 land-cover data for 2024, plus matching ANADEM 30 m terrain tiles covering the full study area. Tide influence: 6.0 m. Sea-level rise: 0.3 mm/year. Soil/suitability layer disabled; migration fallback enabled.
What the check tells us
The two engines produced byte-identical final GeoTIFFs and identical annual trajectories. This verifies implementation equivalence and gives a transparent performance reference; it is not an ecological validation or a universal hardware limit.
Host used for the reference measurement: Intel Core i3-1215U with 19.7 GiB total RAM. Scaling the recorded 78.007-million-cell footprint to 80 million cells gives a first-order estimate of 6.1 GiB for the active state after a step. The continuous step also creates a temporary previous-state snapshot of about 0.7 GiB, so a practical planning figure is roughly 6.8 GiB before application overhead; reserve 8–10 GiB free. These arrays are reused on every step. Therefore, 20 steps do not require 20 times more RAM: memory remains in roughly the same range, while elapsed time grows with the number of steps and annual rasters use disk space. Exporting an annual map can create a short-lived raster buffer, but those buffers are released rather than accumulated. Continuous is faster when the active grid fits in RAM because it avoids persistent block I/O. The 10.01 GiB blocks value in the table belongs to the archived reference run, before the input-grid release optimization. The current block runner copies the input arrays to persistent state files, releases the materialized grid while the simulation runs, and reopens the grid only for post-run maps and figures. Blocks are the capacity and recovery choice for domains that cannot fit safely in RAM, including very large macro-tidal Amazon coastal studies. Actual times and memory vary with CPU load, storage, output settings, and raster size.