The GPU-Resident Astronomy Pipeline
A GPU-resident astronomy pipeline keeps every stage of image analysis — load, align, subtract, detect, review — inside GPU memory instead of round-tripping through the CPU.
Loading post…
NVIDIATanya LenzAnnouncement
Scaling Decision Optimization to 100 Million Variables and Beyond with mPDLP in NVIDIA cuOpt
Summary
Supply chain and energy grid planning problems now exceed the memory and time limits of single-GPU solvers. NVIDIA cuOpt introduces mPDLP, a multi-GPU primal-dual hybrid gradient solver that distributes linear programming problems across NVLink-connected GPUs. The algorithm’s hot loop is a sparse matrix-vector multiplication, a memory-bound operation that benefits from the increased bandwidth of multiple GPUs. Min-cut partitioning reduces communication overhead by keeping tightly connected dependencies on the same GPU, enabling up to 6x lower peak memory per GPU. On the benchmark problem zib03 with over 104 million nonzeros, mPDLP achieved a solve rate nearly 10x faster than the previous year, and partners Kinaxis and PSR demonstrated 3.3x and 5x speedups on real-world models.
Context
Large-scale linear programming problems in supply chain and energy grid planning exceed the memory and time limits of single-GPU solvers, creating bottlenecks for business decisions.
Approach / What changed
cuOpt mPDLP is a multi-GPU primal-dual hybrid gradient solver that distributes LP problems across NVLink-connected GPUs using min-cut partitioning to reduce communication overhead, achieving up to 6x lower peak memory per GPU and faster solve times.
Takeaways
mPDLP uses min-cut partitioning to jointly consider consecutive SpMVs, reducing cross-GPU communication compared to 2D partitioning.
On large problems (nnz > 10^7), mPDLP achieves speedups of 1.2x to 2.5x over D-PDLP.
Partners Kinaxis and PSR achieved 3.3x and 5x speedups on real-world models with over 135 million and 185 million variables respectively.
FAQ
How does mPDLP reduce communication overhead?
mPDLP uses min-cut partitioning to keep tightly connected dependencies on the same GPU, minimizing data exchange across consecutive sparse matrix-vector multiplications.
What hardware is required for mPDLP?
mPDLP requires multiple NVLink-connected GPUs, such as NVIDIA H100 or B200, and uses NCCL for communication.
When does mPDLP outperform single-GPU PDLP?
mPDLP delivers speedups only on large problems where communication overhead is negligible compared to compute time; on smaller problems, synchronization costs outweigh parallelism benefits.
Source and freshness
- Source
- developer.nvidia.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
See the idea
Visual explainers
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.