ProbSplat: Efficient Probabilistic Hardware for Gaussian Splatting in 3D Scene Reconstruction

2026-08-13

AI Summary

This work addresses the challenges of high computational complexity, substantial memory footprint, and excessive power consumption in Gaussian splatting–based 3D reconstruction on edge devices. The authors propose ProbSplat, an in-memory computing architecture leveraging floating-gate inverter arrays to store programmable parameters of Gaussian mixture models—specifically their means and variances—enabling efficient log-likelihood computation. This design allows independent and precise tuning of both mean and variance parameters, significantly enhancing the fidelity of probabilistic distribution fitting while drastically reducing computational overhead, memory requirements, and energy consumption. Implemented in 180 nm CMOS technology, the system operates at 50 MHz with an energy cost of only 18 pJ per inference. When processing 500 Gaussian components, it achieves a mean-variance deviation below 2.4% and a reconstruction PSNR of 21.99 dB.

Abstract

This paper presents ProbSplat, a Compute-in-Memory (CIM)-inspired architecture based on programmable and energy efficient floating-gate inverter columns for probabilistic computing. Improving upon our prior work, ProbSplat programs and stores both means and variances of Gaussian mixture components, and evaluates log-likelihood for gaussian splatting during scene reconstruction with high energy efficiency, suitable for robotics and augmented/virtual reality (AR/VR) at the edge. Our proposed scheme enables independent control of both mean and variance via deterministic adjustment of floating-gate MOSFET threshold voltages, increasing the fidelity of hardware to program probability distributions. The design is simulated in 180nm CMOS on 1.8 V at 50 MHz and achieves mean-variance independence with <2.4% deviation during 3-D Gaussian mixture modeling. Compared to conventional digital implementations, ProbSplat significantly reduces compute complexity, memory footprint, and power consumption. The scalable framework consumes 18pJ energy per log-likelihood inference with 4-bit precision while operating for 500 mixture functions in a 3-D GMM. Scene reconstruction with ProbSplat's characteristics gave satisfactory fidelity of 21.99 PSNR (dB) at 8-bit precision.

Problem

• Gaussian Splatting
• 3D Scene Reconstruction
• Probabilistic Computing
• Edge Computing
• Energy Efficiency

Innovation

• Probabilistic Computing
• Compute-in-Memory
• Gaussian Splatting
• Floating-Gate MOSFET
• Energy-Efficient Hardware