SuperSplat Ships Compute-Based WebGPU Rendering and Automatic Streamed LOD

A 35 million splat scene that ran at 13 fps in PlayCanvas’s WebGL 2 renderer now runs at 76 fps under its new WebGPU path. It’s a 5.7x speedup, achieved by moving the entire splat preparation pipeline onto the GPU. That number headlines today’s PlayCanvas Engine v2.19.0 release and a pair of SuperSplat platform upgrades announced by CEO Will Eastcott. It’s a compute-based WebGPU renderer and automatic streamed levels of detail for every uploaded scene.

For those who have been following PlayCanvas‘s GitHub libraries, they’ll know that this is the culmination of a large scale initiative. The work extends the unified gsplat rendering system the engine has been building since 2.13. The renderer change is architectural. Where the WebGL 2 path sorts splats on a CPU worker thread, the new renderer runs culling, projection, and a GPU radix sort entirely in compute shaders, leaving only lightweight vertex and fragment work to rasterize from pre-projected data.

The gains scale with splat count, with parity at 1 million splats on an M4 Max, 2.6x at 10 million, 5.7x at 35 million. Additionally mobile benefits at every size tested. An iPhone 13 Pro Max roughly doubles its frame rate from 1 to 4 million splats, hitting 78 fps where WebGL 2 managed 38. PlayCanvas cites caniuse.com figures putting WebGPU at roughly 85% of end users, and the engine falls back to WebGL 2 automatically with identical visual output.

The second upgrade attacks load times. SuperSplat now automatically processes every upload into Streamed SOG, an extension of the compressed SOG format in which splat-transform decimates a scene into a series of LODs and chunks them into small streamable pieces. The viewer fetches the lowest detail levels first, putting a complete stable image on screen almost immediately, then progressively refines toward a per-device Gaussian budget. The demo scene attached to the announcement, Guy Middleton’s scan of the Lokwelt Freilassing railway museum in Bavaria, carries 24 million Gaussians at its top LOD, more than any phone could hold in memory, yet appears near instantly and sharpens during exploration.

The pieces have been landing in public all week. Splat-transform 2.5.1 shipped the documented streamed SOG production workflow for anyone who wants to generate the format outside the platform, and v2.19.0 also carries the splat shadow and PCSS work merged earlier in the release cycle. Everything involved, engine, SuperSplat editor and viewer, SplatTransform, remains MIT-licensed open source.