# This Is Now 66x Faster Source: https://www.youtube.com/watch?v=vsK4Gb7Eys8 Recap page: https://rapidrecap.app/video/vsK4Gb7Eys8 Generated: 2026-02-10T11:33:21.382+00:00 --- ## Quick Overview The presented high-performance CPU cloth simulation technique achieves significant speedups over previous methods, such as running 66x faster than C-IPC (CPU) and 2.6x faster than PD-BFGS (GPU) in one complex scenario, by employing domain decomposition and solving only the interactions between adjacent, smaller domains rather than the entire global system simultaneously. **Key Points:** - The proposed high-performance CPU cloth simulation achieves a 66x speedup versus C-IPC (CPU) and a 2.6x speedup versus PD-BFGS (GPU) for a complex 'Hip-hop' simulation with 444K Degrees of Freedom (DOFs). - The method uses domain decomposition, illustrated by the 'Dressed Armadillo' example (150K DOFs, 204ms/frame), to solve local interactions rather than the entire global system. - In a 'Folding' test, the technique maintains realism and handles complex self-collisions and stacking behavior, taking 557ms per frame. - A comparison against PD-Coulomb (CPU) showed an 11x speedup for a complex cloth-on-figure interaction, where the new method only solves interactions for the few critical pieces touching, unlike the previous method which solves all pieces. - The technique also successfully handles complex phenomena like the 'Precision Bridge Shuffle' and 'Twisting Test' which caused previous methods to require excessive iterations or fail. - The performance advantage is demonstrated by solving a 6 million DOF curtain simulation in 6.6 seconds, which would take previous methods an 'eternity' or require 10,000 ants to solve in a comparable time. ![Screenshot at 02:57: The 'Hip-hop' simulation demonstrates a 2.6x speedup over PD-BFGS \(GPU\) and a 66x speedup over C-IPC \(CPU\), highlighting the new method's superior performance on a complex, large-scale garment simulation \(444K DOFs, 588ms/frame\).](https://ss.rapidrecap.app/screens/vsK4Gb7Eys8/00-02-57.jpg) **Context:** This presentation showcases a novel CPU-based cloth simulation framework that leverages domain-decomposed projective dynamics to achieve high performance comparable to or exceeding GPU-based methods. The key innovation involves partitioning complex simulations (like cloth draped on dynamic characters or folding objects) into smaller, interacting domains, allowing for parallelized, localized constraint solving, which drastically reduces computational overhead compared to traditional global solvers. ## Detailed Analysis The video details a high-performance CPU cloth simulation technique based on domain-decomposed projective dynamics. The core idea is illustrated through several complex scenarios. In the 'Dressed Armadillo' test (0:00), the simulation runs in 204ms per frame for 150K DOFs. The 'Folding' example (0:38) shows the method handling complex self-collisions and stacking, achieving 557ms per frame for 492K DOFs. The technique is compared favorably against prior work; for instance, it achieves an 11x speedup over the CPU-based PD-Coulomb (2:32) and significantly outperforms PD-IPC (GPU) in a rope-tying scenario (1:14), maintaining realism while running slightly slower (564ms vs 544ms) but demonstrating better physical fidelity in knot formation. The 'Hip-hop' dance sequence (2:56) is a key benchmark, where the new method runs in 588ms/frame, achieving a 66x speedup over C-IPC (CPU) and a 2.6x speedup over PD-BFGS (GPU). The curtain simulation (1:41), involving 6 million DOFs, is solved in 6.6 seconds, contrasting sharply with the implied 'eternity' required by older methods. The method achieves this efficiency by breaking down the large system into smaller domains, solving interactions locally, and using a 'glue' term (Lambda) to enforce continuity between these domains, as shown in the 'Method' section (5:28). ### Dressed Armadillo - 150K DOFs, 164K Elements, 48 Domains, 204ms per frame - Shows initial performance metrics for cloth on character interaction. ### Folding - 492K DOFs, 324K Elements, 64 Domains, 557ms per frame - Demonstrates handling of complex self-collisions and stacking behavior during cloth folding. ### Hip-hop - 444K DOFs, 291K Elements, 96 Domains, 588ms/frame for Ours vs. 734ms for PD-EXP (GPU) - Highlights 66x speedup vs. C-IPC (CPU) and 2.6x speedup vs. PD-BFGS (GPU) on dynamic character motion. ### Comparison to PD-Coulomb (CPU) - 282K DOFs, 187K Elements, 32 Domains, 145ms/frame for PD-Coulomb vs. Ours - Shows 11x speedup on a character carrying a cloth sheet, emphasizing efficiency in handling complex friction. ### Methodology - Illustrates domain decomposition where neighboring domains agree on shared edges, and only critical boundary interactions are solved, simplifying the global system matrix. ### Related Work Comparisons - Shows Constraint Offset methods requiring much larger simulation steps (e.g., 10mm vs. 1mm) to achieve similar results to the new method with a consistent step size (h=0.04s). ![Screenshot at 0:05: Performance metrics for the 'Dressed Armadillo' simulation: 150K DOFs, 164K Elements, 48 Domains, 204ms per frame.](https://ss.rapidrecap.app/screens/vsK4Gb7Eys8/00-00-05.jpg) ![Screenshot at 0:21: The 'Curtain' simulation demonstrates handling 6 Million DOFs, taking 6.6 seconds per frame.](https://ss.rapidrecap.app/screens/vsK4Gb7Eys8/00-00-21.jpg) ![Screenshot at 1:15: Side-by-side comparison of the new technique versus PD-IPC \(GPU\) showing superior knot formation realism in rope dynamics.](https://ss.rapidrecap.app/screens/vsK4Gb7Eys8/00-01-15.jpg) ![Screenshot at 2:32: Comparison showing the new method achieving an 11x speedup over the previous PD-Coulomb \(CPU\) method in a complex cloth-on-figure scenario.](https://ss.rapidrecap.app/screens/vsK4Gb7Eys8/00-02-32.jpg) ![Screenshot at 4:38: Comparison of the new method \(Ours\) vs. PD-EXP \(GPU\) for a walking animation, showing Ours at 1.1s vs. PD-EXP at 734ms, indicating the trade-off for realism/control on the CPU.](https://ss.rapidrecap.app/screens/vsK4Gb7Eys8/00-04-38.jpg)