Technical note

A faster alternative to SPLAT!

If your coverage studies run overnight, this is what changes at 92 ms a map — and, just as usefully, what SPLAT! still does that BALROG does not.

Written for people who already know what an .lrp file is

SPLAT! is a good piece of software. It is free, it is open source, it implements Longley-Rice honestly, it reads SRTM data, it has been maintained for two decades and it has taught a generation of engineers what irregular-terrain propagation looks like. A large share of the people who evaluate BALROG arrive from it. So this page is not an attack on it — it is an answer to one specific question: what would I gain, and what would I give up?

The one thing that actually changes

Both tools compute the same model over the same terrain. The difference is what a single map costs, and therefore what workflow is possible on top of it.

An area coverage study is not one Longley-Rice call — it is millions. On a CPU, one core walks them in sequence. On the reference NTIA implementation, the 5.95 million receiver points of a 0.6° domain take 55 seconds on one core, or 11.4 s sharded across six. BALROG computes the same grid in 92 ms on a consumer RTX 3060, and each further scenario over the same ground — another frequency, another mast height, another reliability quantile — in 51 ms, because the terrain is solved once and not again.

At 55 seconds a map, you plan the runs. You pick three candidate sites because twenty would take the afternoon, you queue the sweep before lunch, and you interpret what comes back. At 51 milliseconds, the map returns inside the working loop: you drag the transmitter and watch the footprint follow it, sweep the band and see where the shadow moves, and ask questions you would not have bothered to queue.

Three things that stop being batch jobs

QuestionAt ~55 s a mapAt 51 ms
Twenty candidate sites, one band18 min~1 s
Coverage as a volume: six receiver altitudes, ground to 1000 m5 min 30 s347 ms
A transmitter moving along a track, recomputed at every nodea batch job~20 positions / s

The second row is the one that changes what you can sell, not just how fast you work. A link to an aircraft or a drone is a stack of altitudes, not a footprint on a map — and a stack is unaffordable when one slice costs the best part of a minute. The third row is the same argument for anything that moves: a convoy, a flight path, a temporary mast.

What BALROG does not do

This is the part of a comparison page that is usually missing, so it goes above the pricing rather than below it.

What carries over unchanged

On correctness, since it is the obvious worry

A rewrite that is 598× faster and quietly wrong is worse than an overnight job. BALROG is checked against the same NTIA/ITS reference implementation that sets its speed baseline — not against SPLAT!, and not against a CPU version of itself. On smooth ground the two agree to 0.000 dB across sixteen distances; on 108 m of relief the worst of sixteen probes is 1.122 dB, with fourteen inside 0.2 dB and a median of −0.01 dB. Two residuals are not yet explained, and they are named. The full validation note is here, including the 15 dB error the first honest comparison exposed and the two defects behind it.

How to find out in an afternoon

Two routes, neither of which requires trusting this page:

  1. Run it in your browser on a real rented GPU, on one of nine sample terrains or on a tile you upload. You will see the engine's own output and its own timings, not an animation.
  2. Take the 90-day evaluation: the full engine for Linux and Windows, the container image, nine SRTM1 tiles and a working scenario already filled in. Offline licence file, no activation server, no telemetry. Repeat every comparison above on your terrain, your machine, and whatever you use today.
⚡ Run it live on a GPU Request the evaluation

Other notes

How the 92 ms was measured What was timed, on what hardware, with which flags — and the baseline we corrected upward. Agreeing with the reference to 1.1 dB The accuracy work in full, defects included.