Scaling Nonlinear Optimization: Many Problems One GPU
Scaling Nonlinear Optimization: Many Problems One GPU
This paper introduces jaxipm, the first GPU-batched solver for generally constrained nonlinear optimization problems. The authors implement the IPOPT algorithm using JAX to eliminate CPU bottlenecks and reduce GPU idle time by enabling heterogeneous iteration fusion and iteration-level batching. Validated on quadrotor nonlinear model predictive control benchmarks, a single GPU running jaxipm achieves between 30.7% and 69.9% of the throughput achieved by 64 IPOPT instances running in parallel on a 64-core CPU. The open-source codebase is available at https://github.com/johnviljoen/jaxipm.
The authors present jaxipm, a JAX-based implementation of the IPOPT algorithm designed to run nonlinear optimization problems directly on GPU hardware, claiming throughput between 30.7% and 69.9% of a 64-core CPU running parallel IPOPT instances on quadrotor benchmarks.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
robots
64
View original evidence
n fusion, and minimizing GPU idle time through iteration-level batching. Our implementation enables solving hundreds of NLPs in parallel on a single GPU. We validate jaxipm on a range of quadrotor nonlinear model predictive control benchmarks, where on a single GPU we achieve between 30.7% and 69.9% of the throughput of 64 IPOPT instances running in parallel on a 64-core CPU. open-source codebase: https://github.com/
Open source S5
- code: not_reported
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
n fusion, and minimizing GPU idle time through iteration-level batching. Our implementation enables solving hundreds of NLPs in parallel on a single GPU. We validate jaxipm on a range of quadrotor nonlinear model predictive control benchmarks, where on a single GPU we achieve between 30.7% and 69.9% of the throughput of 64 IPOPT instances running in parallel on a 64-core CPU. open-source codebase: https://github.com/
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org