// benchmark

Meld vs Arnis

Meld and Arnis turn the same map into the same Minecraft world. Arnis builds it in one go. Meld cuts the area into tiles, builds many at the same time, then stitches them together with no seams. Same PC, same area: Meld finished about 3x faster (up to 3.4x on the bigger area).

That 3x is the build itself. Once you count the one-time map download, a small first build comes out about even. The real point is scale: the bigger the area, the further Meld pulls ahead, all the way up to whole countries and continents as one seamless world.

~3.4x
faster, big area
2m 24s
Meld, 1024 regions
8m 08s
Arnis, same area
99%
seamless surface
// test rig
CPU Intel Core Ultra 9 275HX RAM 32 GB Disk NVMe SSD Area Bucharest, Romania

How the test works

We built the same Bucharest area two ways on the same PC and timed each. Arnis does it as one build. Meld splits it into tiles, called workers, and builds several at once, each using a slice of CPU, called threads. So "8 workers × 4 threads" means 8 tiles building at the same time, 4 threads of CPU each. The map was downloaded first for both, so we time the building, not the internet.

Every run, timed

AreaHow it was builtTimevs Arnis
8192 × 8192
256 regions
Arnis, one build2m 03sbaseline
Meld, 8 workers × 4 threads41s3.0x faster
Meld, 16 workers × 2 threads44s2.8x
Meld, 4 workers × 8 threads45s2.7x
16384 × 16384
1024 regions
Arnis, one build8m 08sbaseline
Meld, 8 workers × 4 threads2m 24s3.4x faster
Meld, 16 workers × 2 threads2m 24s3.4x
Meld, 4 workers × 8 threads2m 45s3.0x

Best split: 8 workers × 4 threads. Lots of small workers beat a few big ones, and the bigger the area the bigger Meld's lead. Times vary by CPU, RAM and disk.

The download counts too (cold start)

The table above is just the building. There is also the map to fetch. Pulling it straight from OpenStreetMap as you go works fine, but it is faster to grab the whole area once, up front, and reuse it. That is what Meld does: it downloads the map once at the start (or you pre-bake it yourself), then builds from that local copy, offline. Arnis re-downloads every run. Here is the same test with the download counted, from an empty cache:

AreaArnis (download + build, every run)Meld, first run (download once + build)Result
256 regions2m 09s2m 14s (1m 32s download + 41s build)about the same
1024 regions8m 03s4m 08s (1m 44s download + 2m 24s build)Meld ~2x

Cold start, measured. On a small area it is about the same, because the download is most of that time. On the bigger area Meld is already about 2x even cold, and the difference grows with size. On a re-run Meld is offline (41s / 2m 24s) while Arnis downloads again. With buildings on, Arnis re-downloads even more each run.

How big is that grab? Only about a few tens of MB of map data (roads plus elevation) for these areas, which is under a second on a 100 MB/s line. So why did it take about a minute and a half? Because the free OpenStreetMap service is slow to answer, not your connection, so a faster line barely helps. The good news: you pay it once, then every build is offline.

The real point: scale

On a small area the times are close. The headline is scale: the bigger the area, the further Meld pulls ahead, all the way up to whole countries and continents as one seamless world. Stretching the two real runs above (build time), bigger areas look roughly like this:

SizeRegionsArnis, one buildMeld 8 × 4Faster
our biggest test1,0248m 08s2m 24s3.4x
5x bigger~5,120~41 min~12 min~3.4x
10x bigger~10,240~1h 21m~23 min~3.5x
20x bigger~20,480~2h 42m~46 min~3.5x

Estimated, not measured (the 5x to 20x rows just extend the two real runs). The newest Arnis can stream to disk too, so it will not run out of memory on a big area, it simply keeps building one piece at a time. Meld spreads the same job across your cores, so the bigger you go, the more that head start adds up, all into one seamless world with no seams and no height cliffs.

Method, honestly

Get started Back to home