Developers Rewrite Box3D Physics Engine in CUDA for 300x GPU Speedup

Fable and I rewrote Box3d entirely in native cuda. It's about 30x faster on my GPU compared to my CPU (depending on the situation).

This was my hardest benchmark. I still had to be there, but soon I won't. Cost me about 200 dollars

I took a conference talk from Dennis (@voxagonlabs), transcribed it, told fable to make the same tools that he spoke about, and then we drove down the time per frame aggressively.

Big credits to both Dennis & Erin for teaching how this stuff works. 80% of the perf ideas were fable porting what Erin did.

I corrected fable's wrongs a lot of the time, but not because I knew it was wrong -- instead, I was skeptical and told it to check using profiling tools

Music in the video is from Cynthoni's newest album, check it out on bandcamp

This really is insane. A lot of this wouldn't be possible without frontier models, but it's also not possible without someone like me. I'm still necessary. Reading GPU gems 3 because I thought the cover was funny is why I can ask the questions.

@voxagonlabs me and gpt 5.5 found a bug that didn't make it deterministic and fixed it

@voxagonlabs It would probably cost 300 more dollars or so to port this entire thing to webgpu and wasm + make a multiplayer game using webrtc. This is a deterministic physics engine, which makes it even more remarkable

I can do this because I have understanding. I understand the landscape. I understand which hills are worth climbing, I understand how high they go

Don't stop understanding!

I’ve never seen that many stacked rigidbodiess in realtime. This is insane.

Box3D just came out, and Yacine ported it to run on GPUs with roughly 30x the performance using Fable.

@timsoret Well. Fable did some porting, left some god awful bugs, which I fixed with gpt "5.5"

@0xgilbert

@yacineMTB joking, it's very cool what you did

but I think we humans will be in the loop for a bit more

these models are too low entropy

@ZoldenGames The rendering is very efficient, I specialized in renderers for a bit, using the hardware as intended. I obviously could push it much more. I'm getting over 144 FPS on 100k :)

@bonzajplc 100% going to put it up. Just need to very carefully review the code and make sure it has decent edge case perf tests and can't rn cause busy with new born. Maybe in a week

If you want to cut your Fable bill, consider using UsePod. A 1 line change to your harness gets you the exact same Anthropic model at 70% off.

Take a look at our price comparison: https://usepod.ai/marketplace/

We do this by creating a marketplace where resellers compete to deliver the lowest model pricing and routing to the cheapest in real time.

@mikubeam I will, I just need to bench it and stuff properly to not waste people's time but I have a new baby so it's gonna be slow

@yacineMTB Fable and I

@Michi07f @voxagonlabs Okay, I checked, fable lied to me and left a bug. But now it's deterministic, modulo some bugs i haven't found yet

@OGALANGLEY stupid benchod model has a bug . waste of money

@StrykerFromFL Thanks to all of the people subscribing to me on twitter, I can justify the token spend. So yes!