Element Zero

youtube.com/watch?v=bwJgif… Proud to present our paper: Augmented Vertex Block Descent This is a new physics solver which can accurately simulate HUGE numbers of objects in real time! Demo code coming later this week. Checkout the project page for details: graphics.cs.utah.edu/research/proje…

I'm vibe coding a new *GPU rigid body physics engine* from scratch for @levelsio 's #vibejam ! 💥 - Supports Rust & Web (via WASM+WebGPU) - 100% written by @anthropic 's Claude Code + Opus 4.6 - Performance optimized by @karpathy / @uditg 's autoresearch: github.com/uditgoenka/aut… - Uses Augmented Vertex Block Descent (AVBD) paper: graphics.cs.utah.edu/research/proje…

Finally got around to finishing the minimal 3D example implementation of AVBD: github.com/savant117/avbd… You can play with the prebuilt web demo here: graphics.cs.utah.edu/research/proje…

@RandyPGaul Beautiful!

New lighting system is going pretty well - using Holographic Radiance Cascades (arxiv.org/pdf/2505.02041)

@VigorFox @GitHubCopilot @AnthropicAI Neat!

Excited to share my latest project: PhysX_AVBD! Forked NVIDIA's PhysX SDK and integrated an experimental position-level Augmented Vertex Block Descent (AVBD) solver—99.9% coded by AI via @GitHubCopilot powered by @AnthropicAI's Claude models. Repo: github.com/VigorFox/PhysX…

🎉 A big congratulations to the #SIGGRAPH2025 Real-Time Live! Best in Show winner, "InfiniteStudio: 4D Volumetric Capture for Film Making and Beyond," and Audience Choice winner "Crazy Fast Physics! Augmented Vertex Block Descent in Action!"👏✨ These awe-inspiring presentations wowed audiences and showcased the very best of real-time innovation. Join us in celebrating their success by sharing your cheers and reactions in the comments below! 🥳👏 s2025.siggraph.org/program/real-t… photo by Matthias von Loebell © 2025 ACM SIGGRAPH

@PierreTerdiman @leonard_coder In that case, sorry for misrepresenting physx! Real number is definitely less than 50x. No sdk available in the short term, so easiest test is probably just reproducing one of my test scenes in peel. Pyramids svene for example was 500k bodies, 10 ms excluding rendering on a 4090

@leonard_coder @EzeroGames FWIW just commenting out the rendering part makes the snippet significantly faster. There could be further gains by tweaking the settings but I can confirm that rendering takes a large part of the frame time.

The comparison to the PhysX snippet is misleading, as the rendering overhead there is massive for that amount of objects. Still, that does not change the conclusion. youtube.com/live/ucRukZM0d…

@leonard_coder @PierreTerdiman That number came from just dividing the frame time of AVBD by that of physx for the 20k scene, which was about 50x. Though that's a good point about the rendering overhead. I guess I incorrectly assumed the rendering overhead was minimal in the physx sample

@PierreTerdiman they claim 50x faster than PhysX, I'm really confused. 50x factor seems really really suspicious to me. Any clue?

If you are attending Siggraph this year, please come see our talk later today on Augmented Vertex Block Descent (AVBD). We will also be showing a live demo at Real Time Live shortly after!

@MichelVandegaer No 3D is not really much more difficult, in terms of the AVBD solver. Mainly it's collision detection that is a hassle. I may work on a minimal 3D version of this when I have more time after Siggraph, if nobody else has already made one by then.

@EzeroGames Cool! Any plans for a 3D version or is that a too big leap, and does it make things far more complex?

Added post stabilization to the AVBD example implementation. This removes one of the tuning parameters, which should help make implementing the method a little easier: github.com/savant117/avbd…

@DevPlus31 Looking at the code, it has written a PBD solver, not AVBD. Be careful with these LLMs!

Did a Augmented Vertex Block Descent demo with Gemini 2.5 Pro worked 1st try. #SIGGRAPH #Gemini #avbd

@twominutepapers Thank you!

@EzeroGames congratulations on the amazing work!

Made it on two minute papers, my work is done! Kidding, many improvements to come 😉

@marcinignac @Cody_J_Bennett Looks great! I'm very curious how crisp the shadows are (tough to tell with this scene). I guess it depends on the probe resolution

@Cody_J_Bennett Cool. Looking forward to it!

Now with specular GI based on Filament BRDF. Real-time via WebGL/modified three.js with TEXTURE_BASE_LEVEL.

@mmm_wafflez Ahh good catch. Will push a fix!

@EzeroGames When you set the gravity to a value close to 0 the scene will become empty and you need to reset it to continue. Just a small bug with the demo I assume.

For anyone who is interested in implementing our new method Augmented Vertex Block Descent (AVBD), you can find the source code for the 2D demo implementation here: github.com/savant117/avbd… You can also play with the demo in your browser! graphics.cs.utah.edu/research/proje…

@SunsetLearn Everything including spatial data structures are built and managed on the GPU. CPU is doing basically nothing!

@EzeroGames Really excited for these GPU-based parallel sim solvers. Did you guys run the benchmarks in the paper with a pure GPU implementation and nothing else, or was this augmented with spatial hierarchy checks for acceleration (BVH or otherwise)?

youtube.com/watch?v=bwJgif… Proud to present our paper: Augmented Vertex Block Descent This is a new physics solver which can accurately simulate HUGE numbers of objects in real time! Demo code coming later this week. Checkout the project page for details: graphics.cs.utah.edu/research/proje…