New in SuperSplat: WebGPU and Streaming Bring Huge Performance Wins
Last month we shipped Software Attribution, Collision Generation and GPU-Powered Histogram. Today we're announcing some of the most significant upgrades to the SuperSplat platform yet — a brand new high-performance WebGPU renderer and automatic, high-quality levels of detail for streamed scenes. Together they make splats load faster and render smoother than ever, from phones right through to high-end desktops.
The scene below (a stunning scan of Lokwelt Freilassing, a Bavarian railway museum) shows both upgrades in action — explore it right here in your browser.
Click or tap anywhere in the scene above to walk there! Splat by Guy Middleton.
🚀 A Turbo-Charged WebGPU Renderer
Today marks the release of version 2.19.0 of the open source PlayCanvas Engine — also rolling out now in the PlayCanvas Editor — and it's a landmark release for Gaussian splatting. It ships a brand new compute-based WebGPU renderer for 3D Gaussian splats.
Instead of sorting splats on a worker thread, the new WebGPU renderer hands the heavy lifting to compute shaders — culling invisible splats, projecting the rest, and sorting them with a fast GPU radix sort. The final rasterization still uses vertex and fragment shaders, but they're now lightweight, drawing from pre-projected data. The payoff is a dramatic jump in frame rates.
How dramatic? Here's the new WebGPU renderer measured against our existing WebGL 2 renderer on desktop:
| Metric | 1M | 2M | 3M | 4M | 6M | 8M | 10M | 15M | 20M | 25M | 30M | 35M |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| WebGL 2 fps | 137.2 | 140.4 | 124.8 | 113.6 | 86.8 | 71.3 | 48.1 | 28.9 | 22.3 | 18.6 | 15.6 | 13.3 |
| WebGPU fps | 138.7 | 153.6 | 135.5 | 146 | 140.3 | 125 | 124.1 | 105.9 | 97.8 | 89.4 | 85 | 75.8 |
| WebGPU speedup | 1× | 1.1× | 1.1× | 1.3× | 1.6× | 1.8× | 2.6× | 3.7× | 4.4× | 4.8× | 5.4× | 5.7× |
And on mobile:
| Metric | 1M | 2M | 3M | 4M |
|---|---|---|---|---|
| WebGL 2 fps | 38.1 | 35.5 | 25.5 | 20.4 |
| WebGPU fps | 77.6 | 73.2 | 52.7 | 42.4 |
| WebGPU speedup | 2× | 2.1× | 2.1× | 2.1× |
WebGPU is now available to roughly 85% of end users according to caniuse.com, and that number keeps climbing. For everyone else, SuperSplat automatically falls back to the WebGL 2 renderer — and because both paths are built to render identically, your scene looks exactly the same either way. No configuration, no compromises.
📡 Automatic Streaming for Near-Instant Load Times
Big scenes deserve to load fast. SuperSplat now automatically processes every upload into a streamed format the moment it lands — no extra steps and no command line. The payoff is near-instant load times and high frame rates, even on older devices.
We call this format Streamed SOG — an extension of our super-compressed SOG (Spatially Ordered Gaussians) format. It's generated by our open source splat-transform tool, which does two things. First, it decimates the scene several times over to produce a series of high-quality levels of detail (LODs). Then it chunks those LODs into small, streamable pieces the viewer can fetch on demand.
Here's the clever part. The SuperSplat Viewer starts by downloading the lowest LODs first, putting a stable, complete image on screen almost immediately. From there it progressively refines the scene, streaming in higher levels of detail according to an overall Gaussian budget tuned to your device — a generous budget on desktop, a leaner one on mobile. You get the best your hardware can handle, automatically, and you start seeing the scene right away instead of watching a loading bar.
The Bavarian railway museum scene at the top of this post is a perfect example: its top level of detail packs a staggering 24 million Gaussians — far more than any phone could hold in memory at once — yet it appears almost instantly and sharpens as you explore.
This pairs perfectly with the new WebGPU renderer: streaming keeps the Gaussian count right-sized for your device, and WebGPU renders what's on screen faster than ever.
💚 Free and Open Source
SuperSplat, SplatTransform and the PlayCanvas Engine are all free and open source under the MIT license. We believe the best tools for 3D on the web should be accessible to everyone.
If you're building a splat-based application, we'd love for you to build it on PlayCanvas. Check out our repos on GitHub:
New to Gaussian splatting on PlayCanvas? Our Gaussian Splatting documentation is the best place to get started.
👂 We Want to Hear from You
What do you think of the new features? What would you like to see next? Come and join us on the PlayCanvas Discord — it's where the world's best splat creators hang out and we'd love to have you there.
See you in there!