Vorion: A RISC-V GPU with Hardware-Accelerated 3D Gaussian Rendering and Training

Abstract

Vorion is a novel RISC-V GPGPU prototype designed to provide dedicated hardware acceleration for the computationally intensive 3D Gaussian Splatting (3DGS) technique. It features a scalable architecture utilizing z-tiling and a Gaussian/pixel-centric hybrid dataflow, requiring minimal changes to existing rasterizers. A minimal TSMC 16nm prototype achieved 19 FPS for rendering, while a scaled design reached 38.6 iterations/s for 3DGS training.

Report

Key Highlights

• First RISC-V GPGPU with 3DGS Acceleration: Vorion is the inaugural general-purpose GPU prototype built on RISC-V architecture that incorporates dedicated hardware for accelerating 3D Gaussian Splatting (3DGS) rendering and training.
• Addresses Computational Bottleneck: The design targets the massive computation required by 3DGS, aiming to make real-time neural rendering and 4D reconstruction feasible on edge devices and workstations.
• High Performance Prototyping: A minimal system prototype fabricated in TSMC 16nm FinFET technology achieved 19 FPS (Frames Per Second) for rendering.
• Strong Training Capability: A scaled design demonstrated 38.6 iterations per second for 3DGS training workloads.
• Fixed-Function Optimization: The architecture capitalizes on the fixed-function nature of 3DGS, suggesting its integration into the graphics pipeline of next-generation GPUs.

Technical Details

• Base Architecture: GPGPU prototype based on the RISC-V instruction set.
• Core Configuration (Minimal Prototype): 8 SIMT (Single Instruction, Multiple Threads) cores combined with 2 dedicated Gaussian rasterizers.
• Process Technology: Prototyped using TSMC 16nm FinFET technology.
• Architectural Innovations: Features a scalable architecture with minimal modifications required to existing traditional rasterizers.
• Parallelism Mechanism: Implements z-tiling to significantly increase processing parallelism.
• Dataflow Strategy: Employs a novel Gaussian/pixel-centric hybrid dataflow to efficiently handle the 3DGS workflow.
• Scaled Training Performance: The scaled design utilizes 16 Gaussian rasterizers to achieve its peak training iteration rate.

Implications

• RISC-V Ecosystem Advancement: Vorion significantly boosts the credibility and capability of the RISC-V ecosystem in the high-performance computing and graphics domain, demonstrating that RISC-V can successfully host complex, state-of-the-art neural rendering accelerators.
• Enabling Edge AI and Real-Time 4D: By providing hardware acceleration for 3DGS, Vorion enables the deployment of complex neural rendering applications, such as volumetric video (4D) capture and real-time interactive 3D scene generation, on resource-constrained edge devices.
• Future GPU Pipeline Design: This work sets a precedent for integrating emerging, computationally heavy neural rendering algorithms directly into the fixed-function hardware pipeline of future commercial GPUs, moving beyond traditional rasterization pipelines.
• Energy Efficiency: Optimizing 3DGS processing into dedicated hardware, rather than relying solely on general-purpose shader cores, suggests significant gains in performance per watt, crucial for mobile and edge computing applications.