POSTER
Boyuan Tian, Siran Xianyu, Sushant Kondguli, Sarita Adve, Minjia Zhang
3D Gaussian Splatting (3D-GS) has become prominent in spatial computing by enabling photorealistic rendering at interactive rates, but its memory demand is difficult to sustain on resource-constrained systems, where GiB-scale models and large dynamic intermediate states drive transient peak memory to several times the model size and can intensify contention with co-running workloads.
To mitigate the loss of effective system capacity caused by such transient peaks, we present GSM, a peak-memory optimization framework for 3D-GS rendering across a wide range of resolutions. GSM is guided by the insight that minor compute overheads can be traded for substantial peak-memory reduction, with part of the cost recovered through smaller downstream working sets. It therefore holistically co-designs memory management across key rendering stages to reduce large transient intermediate states arising from overprovisioning, over-retention, and over-materialization.
We evaluate GSM across two hardware platforms, two rendering systems, four resolutions, and 13 scenes from three datasets. GSM reduces peak runtime memory by up to 4.6x while preserving rendering quality within 0.1 dB PSNR and incurring negligible performance overhead, thereby improving deployability on resource-constrained platforms and increasing memory headroom for co-running workloads.