When to use which tool

Not in the table: EnergyPlus, a different job ^{(footnote 4)}, and SimScale, a general cloud CFD platform ^{(footnote 13)}.

Illustration: the same building in three ways of storing shape. Cell size is exaggerated to show the effect. Sources: infrared.city internal validation report (Oct 2026, available on request), and the tools' documentation.

Most teams use infrared.city to decide, and validate only the cases that need it.

1. infrared.city, measured. Speed basis: staging benchmark of 24 Oct 2026 (api-test.infrared.city, SDK 1.0.0): one 512 × 512 m tile (Vienna, 2,931 buildings; the 500 × 500 m block fits one 512 m tile), 15 July 08 to 18 h, ground UTCI, warm median 3.0 s end to end (4 runs, buildings already fetched; the first run took 14.2 s with the building fetch). Facades: the same block, solar radiation on facades and roofs, 581,334 surface triangles, 6.4 s warm (9.6 s first run); the SDK has no facade UTCI. Large runs, production cloud, 23 Oct 2026: UTCI for one summer day, 08 to 18 h, hourly, 1 m grid, 400 km² of Berlin in 44 s (four 100 km² runs in parallel); 1 km² in about 5 s on reuse, 17 s on a first run. The engine splits any area into tiles and runs them in parallel, so area is not fixed; whole-city runs (about 1,700 km² at 1 m) are a statement by infrared.city, 24 Oct 2026, and our largest timed run is 400 km². Source: infrared.city internal validation report (Oct 2026), available on request. The facade benchmark covers one 500 m block only. Hourly live layer: statement by infrared.city, 24 Oct 2026, and the Products page for the climate live layer. Back to table
4. EnergyPlus is not in the table: building energy is a different job. infrared.city screens many buildings or variants; EnergyPlus goes deep on one building. It runs one building model at a time with thermal zones (energyplus.net, US Department of Energy), and districts need batch runs. About 2 min per annual run for a 15-zone office: 124 s on a 2008 desktop, version 2.2.0 (LBNL run-time report), so current runs may differ. A 2025 preprint (arXiv 2511.19463) ran about 25,000 buildings in under 30 min on a large compute cluster. Free, BSD-like licence (GitHub README). Back to table
6. Statement by infrared.city, 24 Oct 2026: large jobs run across many servers in parallel. Back to table
7. PALM-4U: Maronga et al. (2020), Geosci. Model Dev. 13:1335. LES (optional RANS), building surface and indoor model, comfort indices (PT, UTCI, PET). Cartesian cells, each fully fluid or fully obstacle; studies use 1 to 24 m. Whole city: Maronga et al., ICUC10 (2018), conference abstract, "Building-resolving large-eddy simulations for entire Berlin" (elib.dlr.de/121678): "PALM-4U is able to compute entire city environments like Berlin (about 1 700 km²)" at 10 m, with a nested 1 km² domain at 1 m; computing cost not reported. Largest 2 m test in the 2020 paper: 1.8 × 1.8 km on 1,296 CPU cores. Cost: CoSynHealth project report (project 1443, 2025 to 2026, Central Altona, Hamburg): "Single 2 m simulations ended up consuming a total of around 2000 node hours (NH) for 24h of simulation"; domain size and node type are not stated. Wall time, derived: Krüger, Burmeister and Winkler (2023), "PALM-4U Handbuch für die Praxis", ProPolis, table on p. 26 (PDF), Hamburg Oberbillwerder wind comfort: child domain 400 × 450 × 50 m at 1 m (0.18 km²) nested in a 480 × 480 × 50 m parent at 5 m, no spin-up, 6 h simulated per run, 8 wind directions, HPC Levante with 300 cores, "Rechenzeit" 39 h in total (about 5 h per direction). Arithmetic: 39 h ÷ 8 ≈ 5 h per 6 simulated hours, so 5 h × 24 / 6 ≈ 20 h per simulated day per wind direction. This is a neutral-stratification wind-comfort setup, not a thermal run, and the handbook's "Rechenzeit" is compute time, not stated to be wall-clock. It assumes the time scales linearly with the simulated hours. Built for HPC; small domains can run on workstations. GPL v3 for the model system. Developed at Leibniz University Hannover. Back to table
8. ENVI-met: 3D wind and turbulence fields with plants, soil and walls; grid typically 1 to 10 m, time step 1 to 5 s (help.envi-met.com). Typical model sizes 50 × 50 to 500 × 500 cells (envi-met.com). Windows 10 or 11, 64 bit; business edition uses all CPU cores. Subscription licence, no public prices (envi-met.com/pricing). Run time: measured about 65 h per simulated day for a 0.06 km² block at 1 m on version 5.9 (Sinsel et al. 2026, below). The vendor says version 6.0 "Simulations run up to 10x faster now" (envi-met.info); we have not tested 6.0. The per-km² figure in the table is an extrapolation; see footnote 14. The 65 h comes from Sinsel et al. (2026), "Comparative validation of ENVI-met versions 5.6 and 5.9…", Atmospheric Environment: X 31, 100479, doi:10.1016/j.aeaoa.2026.100479 (CC BY 4.0): 340 × 171 × 35 cells of 1 × 1 × 2.5 m, 5 nesting grids, 168 h simulated, about 1,383 h on 5.6 and about 455 h on 5.9, which is about 65 h per simulated day on 5.9, on one AMD Threadripper Pro 5965WX desktop (24 cores, 512 GB). The authors call it a rough estimate. The paper shows about 3 times faster from 5.6 to 5.9; the 10 times for 6.0 is the vendor's claim and is not in the paper. Salvati and Kolokotroni (IBPSA BS2019, 500 × 300 m at 2 m, version 4.3.2): about 38 to 120 h. Back to table
9. SOLWEIG: a 2.5D model on height rasters of one resolution (UMEP manual, umep-docs.readthedocs.io). Full day: Zonato et al. (2026), Geosci. Model Dev. 19:7389 (GLIDE-SOL, SOLWEIG 2022a based), Dortmund, 3129 × 3113 pixels at 2 m (about 39 km²), 24 h simulated: "Using one NVIDIA A100 (40 GB) GPU, the 24 h simulation completed in 254 s", 1,285 s on CPU of the same node. Kamath et al. (2026), SOLWEIG-GPU, JOSS 11(118):9535, Table 1 (sky view factor and UTCI, mean of 4 to 5 runs, 1,000 × 1,000 pixels): CPU 1,187 s (Intel i7-10700), GPU 47 s (NVIDIA A6000); wall preparation excluded; pixel size and period not stated. Wall surface temperature: GMD 19:7389. Runs as the Python package solweig-gpu (PyTorch, CPU or GPU) or in QGIS (UMEP). Grids above 4 million pixels should be tiled. GPL-3.0. Back to table
12. Orbital Stack (RWDI), from the vendor's own pages only, checked 24 Oct 2026. Three levels. AI (since 28 Sep 2022): "Get wind and climate estimations in seconds" (orbitalstack.com/ai-powered-wind-simulations); about $100 for an AI wind run and $150 to $200 for AI thermal comfort. Rapid CFD, self-serve: "The CFD Engine provides a rapid, RANS based CFD evaluation" (2023 onboarding roadmap, help.orbitalstack.com) and "All results should be available within 2-3 business days" (help.orbitalstack.com/add-new-scenario/); about $350 for wind and $500 to $700 for thermal comfort per run. Detailed CFD and LES: "(available only through RWDI consulting engagements – simulated and delivered in Orbital Stack)" (orbitalstack.com/purpose-built-cfd). Size: "Buildings and structures located within a 300m radius of the study building" and "The total surface area of all structures and ground surfaces combined must be less than 3.5 km²" (help.orbitalstack.com/preparing-the-geometry-guide/). Analyses: wind comfort, thermal comfort, solar, airflow and cladding wind loading (orbitalstack.com/climate-analysis-overview/). Geometry: "All files must be exported from your CAD software in STL file format." Runs in a browser app. No public API or SDK was found on the vendor pages, and no live or hourly product; both are absence claims. No area or hardware is stated for the run times. Prices are the vendor list prices on its pricing page. Back to table
13. SimScale is not in the table: a general cloud CFD platform with an urban module (wind comfort, thermal comfort including UTCI, solar). Its lattice-Boltzmann runs take hours (vendor: "competition-stage results come back in hours", simscale.com/simulations/urban-microclimate/); AI surrogates are shown for other physics, not for urban. Back to table
14. Speed row and chart: a typical perceived heat (UTCI) calculation, as a rough estimate per km² and per simulated day. It assumes the run time scales linearly with area and simulated hours; real scaling is often worse than linear for 3D models, so large extrapolations are the least certain. Source values and arithmetic. infrared.city: measured, about 5 s for 1 km² at 1 m, warm, 08 to 18 h hourly (not scaled to 24 h); also one 512 × 512 m tile in 3.0 s. Larger areas get faster per km² because tiles run in parallel: 100 km² in 37 s is 0.37 s per km², and 400 km² in 44 s is 0.11 s per km² (about 0.1 s). SOLWEIG (Zonato et al. 2026, GLIDE-SOL; the pipeline computes UTCI, per the abstract, and the benchmark times the full pipeline): 254 s on one GPU and 1,285 s on CPU for about 39 km², 24 h, 2 m; 254 / 39 = 6.5 s and 1,285 / 39 = 33 s per km². ENVI-met: extrapolated linearly from a published 168 h run of ENVI-met 5.9 on a 0.058 km² site at 1 m (455 h on a 24-core, 512 GB workstation; Sinsel et al. 2026, Atmospheric Environment: X 31, 100479): 65 h per simulated day, so about 1,120 h per km² per day, which is about 45 days. That is a ×17 extrapolation of one run, so read it as a size, not a measurement. Real run time depends on hardware, domain and settings. A 1 km² domain at 1 m is far above ENVI-met's documented large-model rule of thumb (250 × 250 × 30 cells) and usually needs tiling. The vendor says version 6.0 runs "up to 10x faster"; we have not measured it. At that best case one 0.06 km² block at 1 m would take about 6.5 h. The run computes a full microclimate and writes UTCI and PET for a standard person (ENVI-met help centre, "Thermal Comfort", Sep 2025 ↗). BIO-met adds personal settings and PMV, SET and dynamic comfort. PALM-4U (handbook): measured 39 h in total for 8 wind directions of 6 simulated hours each on 0.18 km² at 1 m with 300 cores, so about 5 h per direction and about 20 h per simulated day per direction; extrapolated to about 111 h per km², which is 4 to 5 days. This is a wind-comfort run, not UTCI. Orbital Stack: no area-time figure is published, so no time is extrapolated. For size only: a site has a 300 m context radius, so π × 0.3² = 0.28 km² (about 0.3 km²), which is about 3 to 4 sites per km². This size is our estimate from the vendor's 300 m radius, not a vendor figure. Back to table
15. SOLWEIG described as "a widely used urban radiation model for spatially resolving pedestrian radiative environments": Kang, Lindberg, Bridoux, Wallenberg and Kim, "Development of SOLWEIG-LEB v1.0: An urban radiation model integrating LAD-based canopy shortwave attenuation and leaf energy balance", EGUsphere preprint egusphere-2026-6123 (not peer reviewed; abstract quoted from the raw page, 24 Oct 2026). Same points, weather and 3D scene for both models; in-sample sites are lighter in the chart. Back to table