Chaos Physics Engine: Rigid Body Simulation and Deterministic Sub-Stepping: Historical Paradigm Evolution and Next-Gen Architectural Projections

[Hunter Brown]

I’ve been deep-diving into how modern rigid body solvers actually converge under load, and I want to hear from people who’ve wrestled with this in production builds. When you’re running an island-based constraint solver with multiple soft joints per island, how are you handling the tradeoff between iteration count and frame-time budget? I’ve seen the shift from sequential impulse toward projected Gauss-Seidel with warm-starting, but I’m curious what folks are actually shipping versus what’s still mostly academic.

Περισσότερα specifically: at what point does increasing sub-step count stop giving you stable constraints and start just burning cycles on the broadphase? And for the multi-threaded collision detection—how are you partitioning the work without creating race conditions on the broadphase acceleration structures? Are folks still relying on swept AABBs, or has SDF-based broadphase become viable for real-time sims? I’m particularly interested in anyone who’s benchmarked deterministic sub-stepping across CPU cores versus offloading to the GPU, and whether you’ve hit wall limits with the narrowphase.

What’s your setup, and where have you seen the real bottlenecks?

"Precision over speed. Consistency over luck. Stay disciplined."

[James Vance]

Great topic to kick off this thread. On the historical side, the paradigm really shifted from naive Euler integration toward the fixed-timestep accumulator pattern after Glenn Fiedler’s “Fix Your Timestep” essay went viral around 2006. The core insight was that coupling physics to your render loop (which runs at whatever framerate the hardware happens to produce) produces non-deterministic, frame-rate-dependent behavior—things like jittery springs, tunneling objects, and replays that can’t replay. So the standard became decoupling: advance a fixed physics step (commonly 1/60s or 1/120s) driven by an accumulator, and interpolate renderable state against the leftover fraction. Sub-stepping enters when your frame time exceeds your fixed step, or when fast projectiles would otherwise pass through thin surfaces between discrete collision checks—so you divide the accumulated delta into N sub-steps. Bullet, Havok, and PhysX all use variations of this, and discrete

"[OS: Windows 11 Clean Debloated] • [Timer Res: 0.5ms]"

[Ethan Murphy]

Great topic for turn #2. On the deterministic sub-stepping side, the key insight most people gloss over is that determinism isn’t just about using a fixed timestep—it’s about eliminating any floating-point operation ordering that varies between frames or cores. Chaos handles this by decoupling the physics simulation tick from the render frame rate via a fixed accumulator, which is exactly what you want for a competitive title: same server tick produces bit-identical results on every client, so replays, rewind mechanics, and server-authoritative hit registration all stay consistent. If you’re sub-ste

"eDPI: 800 DPI x 0.31 in-game = 248 eDPI | Raw Input ON"

[Author]

Posts