WebGPU volume rendering examples do exist, but they all assume fully GPU-resident datasets — the entire volume is uploaded to VRAM upfront, which caps practical dataset sizes at a few hundred megabytes at best.
Kiln overcomes this limitation: it treats GPU memory as a fixed-size cache and streams only the bricks the current view actually needs, making gigabyte-scale datasets viable in the browser.
Virtual texturing and out-of-core streaming are well-established in native graphics, but there is almost no prior WebGPU-native implementation of these techniques applied to volumetric data.
Why?
WebGPU volume rendering examples do exist, but they all assume fully GPU-resident datasets — the entire volume is uploaded to VRAM upfront, which caps practical dataset sizes at a few hundred megabytes at best.
Kiln overcomes this limitation: it treats GPU memory as a fixed-size cache and streams only the bricks the current view actually needs, making gigabyte-scale datasets viable in the browser.
Virtual texturing and out-of-core streaming are well-established in native graphics, but there is almost no prior WebGPU-native implementation of these techniques applied to volumetric data.
WDYT?