The Performance panel✎
What drawing the place costs: frame times, the last frame's counts, and a profile of each stage.
Reviewed
The Performance panel, in the Panels menu beside Rendering, shows what drawing the place costs. It updates twice a second while it is open.
The graph
The last 240 frames' times, with lines at 16.7 ms and 33.3 ms: a frame at 60 and at 30 a second. Where the browser times the GPU, each frame splits into its passes: shadows, the place and effects. Otherwise the graph shows the page's own time.
The last frame
| Row | Counts |
|---|---|
| Draw calls | Batches sent to the GPU |
| Instances | Copies those calls drew |
| Solid parts drawn | How many of the solid parts culling left; All with culling off |
| Triangles | Triangles drawn |
| Beams and Particles | How many were drawn |
| Page | The page's own time for the frame |
| GPU, shadows, place and effects | Each pass's time, where the browser times them. WebGL2 does not |
| Assets | How many of the place's assets are loaded, and any that failed or need signing in |
Profile
Profile this view measures the frame in about a minute, on any device, including ones whose browser cannot time the GPU, as Safari on an iPad cannot. Keep the view still while it runs; the view is put back as it was when it ends.
| Part | What it measures |
|---|---|
| On screen | The frame rate the view gets, at the display's pace |
| One frame | One frame drawn and waited for until the GPU has finished it, and the page's own work for it: effects, culling and the rest. Where the GPU has timers, every row also gives its own time, which leaves out waiting for the screen |
| Stages | Each stage of the frame, the GPU waited for at the end of each: shadows, glow, the parts, post effects and the rest. Their median, and the time 95 frames in 100 come in under. Where the GPU has timers, the same stages by its own clock, which leaves out the waits; the sun's maps are drawn every frame here, as when the camera moves. When the sky maps are drawn again, the thin parts taken out of the near ones and the lighting cells are stages of their own |
| Base | With every feature off: the frame cleared and nothing drawn, the parts' depth alone (WebGPU), and the parts shaded. Apart, they say whether the cost is the screen's pixels, the parts or their shading |
| Light and Shadows | Each switch of those groups turned off alone, and what the frame saves without it |
| Steps | The frame with the Rendering panel's steps added one at a time, of the switches the view has on; each step's cost is how much slower it got |
| Resolution | The frame at all, half and a quarter of the pixels. Much faster with fewer means filling pixels is the cost |
| Antialiasing | One frame in each antialiasing this device draws, the view built again for each, so they compare. Multisampling also with the parts shaded and depth-tested in one draw |
| Shader precision | With WebGPU where the device has half precision: one frame with the colour arithmetic in half, and in full, the view built again for each |
| Depth path | With WebGPU: one frame with the parts' depth drawn first, and with them shaded and depth-tested in one draw, as a tile-based GPU (Apple's) wants |
| Camera moving | The camera orbited, panned and zoomed as a drag and a pinch move it, and moved by each touch stick, for one second each from where it was, at the same speed on every device: the frame, the GPU's stages, and how often the reflection cube and the near sky maps were drawn again |
| Quality levels | Each level, 1 to 21, with its own render scale and antialiasing: its frame, its stages (by the GPU's timers where it has them), and what it drew at: the canvas, antialiasing, the sun's map and reach, the glow's size, draw calls, and the switches it adds. The view goes back to its own level after |
| Device | The renderer, the GPU's name, the canvas size, the level and antialiasing the view was at, and the camera, to take the next profile from the same spot |
Info
Copy report copies the whole profile as text, to paste into an issue or a message.
Benchmark
Roswaal Benchmark 4 draws fifteen test scenes that Roswaal makes for itself, the same on every device, and times each on the GPU and the CPU, with the GPU kept busy, two frames in flight, so a GPU that slows itself when lightly worked runs at full speed. Start it from Settings › Scene view or the Profile section here: Current settings runs the view as it is set, about 2 minutes; Full sweep runs every quality level, about 40 minutes. The editor is hidden while it runs. The Roswaal logo, or Esc, stops it and brings the editor back as it was.
| Station | What it measures |
|---|---|
| Empty | What every frame costs |
| Materials | Every material in three colours, at three Reflectances |
| Textures | Kenney's prototype textures as Textures on every face, and a tiled floor |
| Shapes | 1,200 parts of every shape |
| Meshes | 96 tanks of MayLizbeth's, 1,788 MeshParts, their skirts and tracks DoubleSided |
| Sky light | Closed rooms with walls 1 to 8 studs thick, high roofs, struts and rails |
| Sun shadows | Towers to 400 studs and a field of thin poles |
| Lights | 10, 20, 50 and 100 lights in one area |
| See-through | Glass stacks, Transparency in steps, ForceField |
| Glow | Neon in colours and sizes, behind glass and through a seam; CrackedLava |
| Effects | 100 Beams; 1,000, 10,000 and 20,000 particles |
| Highlights | Both depth modes, a Model, an Adornee, a transparent part |
| Air and post | Atmosphere, fog, Bloom, ColorCorrection and Blur |
| Dense | 19,600 parts in a city grid |
| Moving | The Dense grid with each control's motion |
Each station's score is the frames a second its GPU time allows, beside the rate its waited frame allows; the run's is their geometric mean. Scores compare within one benchmark version. The tanks are Teensy Tiny Tankery models by MayLizbeth ("extraclear"), CC BY 4.0. A run draws with the Rendering panel's switches as they are, and its report names any that are off. A/B a switch, in Settings, measures every station with one switch on, off, on and off, back to back, so a warming device slows both alike, and says what the switch saves or costs at each. Copy report gives the text; Download results saves the whole run as JSON.
Compare with last run, at the end of a run, or Compare two results in Settings, sets two runs side by side. A warm tablet is slower everywhere, so the change across all the stations (their median) is taken as the run's drift and taken out of each station's; a change inside the noise, how far the stations scatter, is called within noise. Warm pass, in Settings, measures every station again after the whole first pass, when the device has warmed up, and shows how much slower each is warm; scores and comparisons stay the first pass's. It takes twice as long.
Finding a spike
Fly round and watch the graph for frames over the 33 ms line. Then, in Rendering, turn switches off one at a time, or freeze the shadows, and see which brings the time down. Mesh Debug names the meshes to simplify. To watch the place through a panel, pick See-through in its ⋯ menu.
Last reviewed 20 November 2026 by @neopolitans
To verify: The panel is still in progress: its page changes with it.