Performance

Measured with npm run bench on 2026-09-06: Apple M1 Pro, 32 GB RAM, headless Chrome, a 1280x800 canvas, the Blue Marble and Landsat imagery, atmosphere and star field as the base, and the camera orbiting one degree of heading per frame for 120 frames. "Draw time" is WorldWind's own render time per frame (BEFORE_REDRAW to AFTER_REDRAW); "Setup" is the time to build the objects and draw the first frame that includes them; "Heap growth" is the JavaScript heap after setup minus before it, each measured after a forced garbage collection. Imagery tiles are loaded from NASA's servers before measuring, so the numbers describe rendering, not the network.

Every object is a real WorldWind renderable created through the kit's factories; nothing is batched or simplified by the library, so these are WorldWind's numbers as reached through worldwind-kit.

GPU

Renderer: ANGLE (Apple, ANGLE Metal Renderer: Apple M1 Pro, Unspecified Version)

Scenario	Objects	Setup	Draw time avg	Draw time p95	Frames/s	Heap growth
baseline	0	0 ms	2.1 ms	3.3 ms	60	0.2 MB
placemarks-1k	1,000	12 ms	2.9 ms	3.5 ms	60	1.9 MB
placemarks-10k	10,000	50 ms	14 ms	18.6 ms	60	15.1 MB
placemarks-50k	50,000	204 ms	69.1 ms	73.5 ms	14	74.3 MB
clustered-10k	10,000	48 ms	5.2 ms	6.1 ms	60	3.8 MB
clustered-50k	50,000	56 ms	5 ms	5.5 ms	60	5.8 MB
geojson-10k	10,500	154 ms	23.5 ms	25.3 ms	42	18.9 MB
paths-500	500	14 ms	2.9 ms	3.5 ms	60	2 MB

Software (SwiftShader)

What a CI runner or a machine without a usable GPU sees. Renderer: ANGLE (Google, Vulkan 1.3.0 (SwiftShader Device (LLVM 10.0.0) (0x0000C0DE)), SwiftShader driver)

Scenario	Objects	Setup	Draw time avg	Draw time p95	Frames/s	Heap growth
baseline	0	1 ms	1.3 ms	1.5 ms	22	0.2 MB
placemarks-1k	1,000	18 ms	2.5 ms	2.7 ms	19	1.9 MB
placemarks-10k	10,000	63 ms	110.4 ms	127.3 ms	9	15.1 MB
placemarks-50k	50,000	219 ms	407.8 ms	454.2 ms	2	74.3 MB
clustered-10k	10,000	50 ms	55.6 ms	67.3 ms	16	3.8 MB
clustered-50k	50,000	63 ms	56.2 ms	67.3 ms	16	5.8 MB
geojson-10k	10,500	447 ms	160.8 ms	211 ms	6	19 MB
paths-500	500	24 ms	2.2 ms	2.5 ms	19	2 MB

Reading the numbers

• Placemarks are the expensive object: each is a textured quad with its own picking colour, and WorldWind sorts and draws them one by one. Ten thousand still hit 60 frames per second on a laptop GPU; fifty thousand do not. The clustered rows draw the same points through ClusterLayer: only a few hundred markers reach the GPU at any zoom, so fifty thousand cost about as much as one thousand. The software renderer shows where the cost goes.
• GeoJSON polygons are surface shapes: tessellated onto the terrain once, then cheap per frame. The points in that scenario are placemarks and dominate its draw time.
• Paths are tessellated once and cost almost nothing per frame.
• Heap growth is roughly proportional to object count and is released by globe.destroy() or by removing the layer.

To reproduce: npm run build -w bench && npm run bench. The scenarios live in examples/bench/src/main.ts.