Home / ForgeGIS

ForgeGIS™

GPU geospatial compute for Java, with a native MCP server.

ForgeGIS runs terrain, hydrology, visibility, spectral, zonal and vector operations on the GPU, along with resampling and reprojection. Data stays on the GPU between stages, so a pipeline does not copy back to the CPU after every step, and every operation has a CPU fallback. It ships as Maven artifacts, with no GDAL at run time and no native binaries of its own.

Request an evaluation Technical brief (PDF) ↓
How it runs

One engine, three ways in.

From Java

Add the Maven artifacts and call operations from your own code. Rasters stay on the GPU until you read a result back.

From an agent, over MCP

The built-in MCP server gives an AI agent the same operations. A request ForgeGIS can’t run comes back as a typed refusal that says why, so the agent can correct it.

As pipelines

Chain operations into a graph and submit it as one run_dag call; intermediate results stay on the GPU from stage to stage.

The ForgeGIS Studio canvas in dark mode: a dataset catalog on the left, and on the canvas a 3600 by 3600 elevation source node above a HillshadeOp node showing a grey hillshade thumbnail
A two-node pipeline in ForgeGIS Studio: an elevation dataset feeding a hillshade.
The benchmark

Checked against GDAL, and faster.

ForgeGIS 1.2.6 against GDAL 3.12, warm timings.
15×faster than GDAL, geometric mean across 242 benchmark operations, on rasters from 16×16 to 8192×8192 pixels and an NVIDIA RTX 5070 Laptop GPU.
6×on operations of 1 ms or more, GDAL in process.
1,191 checks, 0 failuresparity checks on ForgeGIS’s output, each within a tolerance stated per operation.
239 of 242 fasteroperations where ForgeGIS finished ahead of GDAL.

3 of 242 operations were slower, all small rasters.

Method

How it was measured.

  • Benchmark run 20261004-094534-191f1b44, written up in the technical brief.
  • GDAL is called in process through its Java bindings, so both sides are timed on the operation, not on starting a program.
  • Warm runs, the median of each block, with the GPU kept in its working state between blocks, as a sustained workload would see it. The timing method was tightened in this release, so earlier published figures are not directly comparable.
  • The geometric mean covers every operation with a GDAL equivalent and both sides timed: no category excluded, sub-millisecond operations included. 884 of the parity checks compare ForgeGIS’s output with GDAL’s or an independent oracle; the rest are consistency and smoke checks.
  • “Warm” means after the operation has run once. The first call after the GPU has gone idle costs more, and can be slower than GDAL for sub-millisecond work.

NVIDIA RTX 5070 Laptop GPU · Windows 11 · GDAL 3.12.1. Results are specific to this setup.

Speedup by category
CategoryGeometric meanOperations (n)
Pipelines81×14
Resampling30×51
Transforms30×13
Vectorizing26×44
Warps24×20
Zonal and focal20×29
Terrain4.3×5
Viewshed3.4×2
Raster I/O3.0×64
The race

Watch it run.

Operation 1 of 61, a hillshade of the Grand Canyon: the ForgeGIS and GDAL results side by side, with times of 0.044 s and 0.176 s beneath
Not the benchmark: 61 operations run once each, side by side, real time. Cop30 DEM contains modified Copernicus DEM data © DLR e.V. 2010–2014, Airbus DS GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA. Sentinel-2 L2A contains modified Copernicus Sentinel data, ESA. Credits.
Technical brief

The method in full.

ForgeGIS Technical Brief

Validation method, benchmark details and the architecture, for engineers evaluating ForgeGIS.

Download PDF