GPU-Accelerated NMPC framework achieves 94% faster solves
New parametric method reuses computations, cutting per-iteration solve times by 94%.
Gondosiswanto and Pulsipher's new framework tackles the computational bottleneck of nonlinear model predictive control (NMPC), which requires solving complex optimal control problems in real time. Traditional GPU-accelerated methods rebuild entire problems from scratch at each step, even when only measurements or references change. The authors introduce a parametric interior-point formulation that preserves the transcribed problem's fixed structure, allowing critical computations like sparse Cholesky factorization to be reused across successive solves.
Tested on distillation column and 2D heated plate benchmarks, the framework achieves over an order-of-magnitude speedup in total NMPC runtime compared to state-of-the-art CPU and GPU configurations. More strikingly, per-iteration solve times on GPU dropped by 94%, dramatically widening the envelope for real-time control applications. The work positions GPU-accelerated NMPC as a practical tool for systems with fast nonlinear dynamics, such as robotics, autonomous vehicles, and industrial process control.
- New parametric interior-point method reuses factorization across solves, avoiding redundant computation.
- Achieves over 10x speedup in total NMPC run time on distillation column and 2D heated plate benchmarks.
- GPU execution reduces per-iteration solve times by up to 94% versus CPU baselines.
Why It Matters
Enables real-time nonlinear control for robotics and industrial processes by slashing computational costs on GPUs.