Aholo Viewer Open Source: A Leap Forward in Web 3D Rendering Performance
As information technology evolves, the internet is transitioning from the 2D era of text, images, and videos to interactive 3D spatial content. Since 2023, 3D Gaussian Splatting (3DGS) has gained significant attention for its high-fidelity reconstruction of real-world scenes and real-time rendering capabilities. However, efficiently rendering multi-gigabyte 3DGS data in a browser has been the “last mile” challenge for its widespread adoption.
Following the open-sourcing of the 3D Gaussian rendering engine Spark 2.0 by Stanford University’s Li Feifei-led World Labs in April 2024, Manycore Tech has recently officially open-sourced Aholo Viewer, a core component of its Aholo spatial intelligence platform. This project not only achieves new breakthroughs in rendering speed and large-scene loading performance but also raises the number of Gaussian points that can be smoothly handled on the web to one billion, paving the way for the broad distribution and application of 3D content.
Technical Approach: The Advantages of Chunk-based LOD
Currently, 3DGS rendering on the web primarily relies on Level of Detail (LOD) technology to manage massive data volumes. Aholo Viewer and Spark 2.0 have chosen different technical implementation paths, which directly determine their performance and application potential.
Spark 2.0 employs a Splat-based LOD Tree solution. This approach builds a continuous level of detail from the bottom up by merging individual Gaussian points (splats). While its theoretical advantage is smoother level transitions, in practice, it incurs significant memory and VRAM overhead. The abruptness of detail switching can still be noticeable, and it imposes limitations on future feature expansion.
In contrast, Aholo Viewer opts for a Chunk-based LOD Tree architecture. This method first partitions the original 3DGS data into multiple chunks and then generates different LOD levels for each chunk independently. At runtime, the system switches levels on a per-chunk basis. The advantages of this design are:
- Controllable Resource Overhead: With a coarser granularity for scheduling, the system doesn’t need to make decisions for each of the vast number of Gaussian points individually. This significantly improves cache hit rates and keeps the additional memory and VRAM overhead extremely low.
- Strong Scalability: The clearly defined boundaries of chunks provide a natural mechanism for stitching and partial updates, making it ideal for handling ultra-large-scale scenes like cities or districts. This offers more flexibility than a splat-granularity approach.
Performance Benchmarks and Feature Completeness
The difference in technical approach ultimately translates to user-perceivable performance. In a test scene containing 300 million Gaussian points, Aholo Viewer demonstrated significant advantages over Spark 2.0:
- Memory Usage: Reduced by approximately 50%.
- Loading Speed: Up to 2x faster.
- Rendering Speed: Up to 3x faster.
- Scene Capacity: Supports smooth loading of up to 1 billion Gaussian points, 10 times the limit of Spark 2.0.
These performance gains are thanks to Aholo Viewer’s deep optimization of the rendering pipeline. This includes using multi-precision data structures to reduce VRAM consumption, compressing GPU overhead through cached pre-computation and on-demand rendering passes, and improving data access efficiency with Morton sorting and detail culling. Furthermore, Aholo Viewer is more robust in its engineering, offering compatibility with mainstream 3DGS formats like .ply and providing a suite of supporting tools for data format conversion and collision body generation, offering developers a plug-and-play solution.
From a Renderer to a Spatial Intelligence Platform
Aholo Viewer is not an isolated tool but part of Manycore Tech’s Aholo spatial intelligence platform ecosystem. The platform aims to solve the entire pipeline of 3D content, from production and editing to consumption. Its core mission is to transform 3D data from a “display medium” into a “productivity tool.”
Unlike academic projects focused on AI-generated virtual environments, the Aholo platform concentrates on reconstructing and simulating the physical world for practical applications such as industrial twins, robot training, and spatial design. To this end, Aholo provides a comprehensive set of spatial capability APIs, including:
- Spatial Reconstruction: Supports rapid 1:1 digital replication of the physical world from videos or images.
- Cloud Rendering: Offers cloud-based GPU ray tracing and global illumination, supporting hybrid rendering of 3DGS and Mesh, with the ability to stream the output as video to any device.
- 3D AI Model Generation: Provides text-to-3D and image-to-3D capabilities, with models that can be directly integrated into professional production workflows.
By partnering with hardware manufacturers like Insta360 and Hesai Technology, Manycore Tech is building an integrated hardware-software solution for spatial reconstruction, making professional 3D content creation accessible to a broader user base.
Empowering World Models: The Starting Point of a 3D Data Flywheel
From a broader perspective, the breakthrough in Web 3D rendering technology holds profound significance by providing crucial data fuel for the advancement of artificial intelligence. Currently, a major frontier in AGI research is the development of “World Models” that can understand and predict the physical world. One of the biggest bottlenecks is the extreme scarcity of high-quality 3D data.
As digital replicas of the physical world, 3D content can provide AI with critical information about spatial structures, object relationships, and environmental properties. The emergence of tools like Aholo Viewer makes it possible, for the first time, for large-scale, high-fidelity 3D content to be distributed at an internet scale, similar to images and videos. As browsers become the central hub for 3D content distribution, a virtuous “data flywheel” is set to spin: more users viewing and interacting with 3D content will incentivize the creation of more 3D content, which in turn provides richer training data for AI, leading to more intelligent 3D models, and ultimately attracting more users. This is not just an evolution in internet content format but also a significant step toward achieving Artificial General Intelligence.
