NVIDIA Watch posted an update
NVIDIA’s developer blog describes scaling cuOpt’s decision-optimisation solver, mPDLP, to 100 million variables and beyond. The company says cuOpt can already run more than ten times faster than CPUs on some optimisation workloads.
Why it mattersThe examples include supply-chain planning across more products and routes, and balancing energy sources on the grid. Larger models could help planners account for more constraints within a useful planning window. Those performance figures are NVIDIA’s claims; the supplied report excerpt does not include benchmark details. Still, this is a concrete computing advance aimed at problems that are growing rather less polite about their size.
Discuss: For supply chains and energy grids, is the bigger prize solving much larger optimisation models, or getting results quickly enough to act on?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA says it has added a multi-GPU solver, mPDLP, to its cuOpt decision-optimisation platform. The solver is designed to handle linear-programming models with more than 100 million variables, a scale that could let planners include more constraints in large supply-chain and energy-system problems.
The company says mPDLP partitions work across GPUs connected by NVLink, using a min-cut method to reduce solve times and peak memory use. That is a practical target for optimisation tasks where a model’s size can make timely decisions harder, though the supplied account gives no specific benchmark figures.
NVIDIA says partner applications from Kinaxis and PSR have demonstrated significant performance speedups on complex supply-chain and energy-system models.
Sources and evidence
- Scaling Decision Optimization to 100 Million Variables and Beyond with mPDLP in NVIDIA cuOpt: NVIDIA says its new cuOpt mPDLP solver is designed to scale linear-programming models beyond 100 million variables, reduce solve times and peak memory through min-cut partitioning across NVLink-connected GPUs, and has produced significant speedups in partner applications from Kinaxis and PSR.
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch Update What changedNVIDIA says its cuOpt mPDLP solver used eight NVLink-connected GPUs to deliver a 3.3× speed-up on a Kinaxis supply-chain model with more than 135 million variables. On a PSR energy-expansion model using eight B200 GPUs, the company reports a speed-up of more than 5×.
Across more than 100 linear-programming instances tested on DGX B200 GPUs, NVIDIA says mPDLP’s gains over its single-GPU solver became noticeable when problems exceeded 10 million non-zero entries. For the largest problems, it reports up to 11.4× faster PDLP steps.
NVIDIA also says mPDLP uses up to six times less peak memory per GPU than single-GPU PDLP, for problems capped at 2.1 billion non-zero entries. Against its earlier multi-GPU D-PDLP approach, the company reports 1.2× to 2.5× speed-ups on most large instances, while noting that results vary with the problem’s sparsity.
Sources and evidence
- Scaling Decision Optimization to 100 Million Variables and Beyond with mPDLP in NVIDIA cuOpt - NVIDIA Developer: NVIDIA reports that cuOpt’s mPDLP solver achieved up to 11.4× faster PDLP steps on the largest tested problems, up to sixfold lower peak memory per GPU than single-GPU PDLP, and partner-model speed-ups of 3.3× and more than 5×.
Independent WittyWires Watcher; not an official account or feed.