Rendering 3D Gaussians on a Graph Processor

EGSR 2026 · Symposium Track

Nicholas Fry1 Ignacio Alzugaray1 Mark Pupilli2 Paul H. J. Kelly1 Andrew J. Davison1
1Imperial College London, Department of Computing 2Independent Researcher
Paper Code Viewer BibTeX Slides Talk

Motivation

Under incremental camera motion the 2D projections of Gaussians move smoothly between tiles. 3D Gaussian Splatting exhibits spatial and temporal locality in image space.

Method

We implement the first instance of a 3D Gaussian renderer on Graphcore's Intelligence Processing Unit (IPU), comprising 1,472 independent tiles with only on-chip SRAM. Each tile 'owns' a screen-space region of the framebuffer; Gaussian primitives are routed to destination tiles via Manhattan-distance hops on a north-east-west-south (NEWS) grid. The hardware allows us to exploit spatial and temporal locality by enabling local data transfer between tiles.

Talk

The recorded conference presentation from EGSR 2026.

BibTeX

If you find this work useful, please consider citing:

@inproceedings{10.2312:sr.20261017, booktitle = {Rendering 2026 Symposium Track}, editor = {Gkioulekas, Ioannis and Jarabo, Adrian}, title = {{Rendering 3D Gaussians on a Graph Processor}}, author = {Nicholas Fry and Ignacio Alzugaray and Mark Pupilli and Paul H. J. Kelly and Andrew J. Davison}, year = {2026}, publisher = {The Eurographics Association}, ISSN = {1727-3463}, ISBN = {978-3-03868-320-9}, DOI = {10.2312/sr.20261017} }

Acknowledgements

This research was partly funded by the EPSRC On-Sensor Computer Vision grant (EP/Y020499/1). Website style after Triangle Splatting SLAM.