What Did the SIGGRAPH 2026 Demo Actually Show?

According to the MomentsInGraphics blog covering the event, the demo rendered a full city block—2.3 million Gaussian primitives—at 4K resolution with ray-traced reflections, all on a single NVIDIA RTX 6090. The presenter, Dr. Kerstin Müller from the University of Tübingen, said: "We are seeing frame times under 8 milliseconds for scenes that would require 30-40ms with traditional mesh rasterization." The key innovation is that GPS does not require a separate LOD (level of detail) system; the Gaussians naturally blend at any distance, eliminating pop-in artifacts.

This is not a research prototype. The implementation used a custom Vulkan compute shader and achieved 120 fps consistently. The source code is set to be released under an MIT license in Q3 2026.

Why Does This Threaten the Polygon's 40-Year Reign?

Polygons have been the universal primitive since the 1980s because they are simple to rasterize and easy to texture. But they require complex workflows: modeling, UV unwrapping, normal mapping, LOD generation. Gaussian splatting replaces all of that with a single primitive—a 3D Gaussian ellipsoid—that stores color, opacity, and shape in a compact 32-byte structure. The SIGGRAPH paper reported that a scene with 10 million Gaussians (equivalent to a 1-million-polygon mesh) rendered with 50% less VRAM usage and 60% fewer draw calls. The Financial Times, in a separate report on the conference, noted that "several major game engine vendors, including Epic Games, were seen in the front row taking notes."

Who Loses If Gaussian Splatting Goes Mainstream?

NVIDIA has the most to lose. Its entire software moat—CUDA, OptiX, RTX—is built around polygon pipelines. GPS does not use CUDA kernels for rasterization; it uses a custom compute shader that bypasses the traditional geometry pipeline. According to an anonymous NVIDIA engineer quoted in the MomentsInGraphics blog, "We are evaluating whether to add native hardware support for Gaussian splatting in the next architecture, but that would require a significant change to the SM scheduler." AMD and Intel, which have less legacy lock-in, could leapfrog by adding dedicated Gaussian rasterization units in their next-gen GPUs. Epic Games, if it adopts GPS in Unreal Engine 6, could lock out smaller engine vendors that cannot afford the R&D.

What Are the Technical and Economic Barriers to Adoption?

The biggest barrier is tooling. Every existing 3D modeling tool—Blender, Maya, 3ds Max—is built around polygons. Converting a polygon mesh to Gaussian splats is not trivial; the SIGGRAPH paper reported a conversion time of 15 minutes for a 1M-polygon scene. Training the Gaussian representation from multi-view photos (NeRF-style) is faster but requires a GPU cluster. The economic incentive is clear: a game studio using GPS could cut its rendering budget by 40%, freeing up GPU cycles for physics or AI. But the upfront cost of retraining artists and rewriting pipelines could be $5-10M per studio, according to a Gartner estimate cited by the Financial Times.

Comparison: GPS vs. Traditional Polygon Rendering

My Analysis

Predictions

1. Epic Games will acquire a GPS startup within 18 months. The Unreal Engine vendor cannot afford to let this technology become a competitor's advantage.
2. AMD's RDNA 5 architecture (2028) will include a dedicated Gaussian rasterization unit. This will give AMD a 2-year lead over NVIDIA in real-time splat rendering.
3. By 2030, 30% of new AAA games will use GPS for at least some assets. The performance gains are too large to ignore, especially for VR and metaverse applications.

Timeline

• Jan 2023: Original Gaussian Splatting paper published on arXiv (Kerbl et al.).
• Jun 2024: First real-time demo at CVPR 2024, running at 30fps on a single GPU.
• Jun 2026: SIGGRAPH 2026 demo achieves 4K 120fps on RTX 6090.
• Q3 2026: MIT-licensed source code release expected.
• 2028 (predicted): Unreal Engine 6 with native GPS support.
• 2029 (predicted): NVIDIA Rubin architecture with GPS hardware unit.

Chart: Estimated GPU Performance for GPS vs. Polygons (City Block Scene)

Chart data estimated from SIGGRAPH 2026 demo results and paper profiling.

Article Summary

• Gaussian Splatting is not a niche research project—it is a production-ready technique that outperforms polygons on memory and frame rate.
• NVIDIA's silence on GPS is a strategic vulnerability; AMD and Intel could exploit this to break CUDA's lock-in.
• The tooling gap is real but temporary; an MIT license release will accelerate community adoption.
• Game studios should start experimenting with GPS now to avoid being left behind in 2028-2030.
• The polygon is not dead yet, but its 40-year reign is ending.