# Nvidia details multi-GPU cuOpt solver with lower memory requirements

_Published Wednesday, October 7, 2026 at 10:35 PM EDT · Infrastructure · Latest · Tier 2 — Notable_

![Nvidia details multi-GPU cuOpt solver with lower memory requirements — Primary](https://developer-blogs.nvidia.com/wp-content/uploads/2026/10/geometric-structure-1.webp)

Nvidia introduced a multi-GPU solver for cuOpt that distributes linear programming problems across NVLink-connected GPUs. These models find an optimum under constraints, supporting tasks such as supply planning and energy-system expansion. Nvidia reports up to 6x lower peak memory use per GPU than its single-GPU solver, with problems capped at 2.1B nonzero entries.

The solver partitions related calculations to keep more work on each GPU and reduce communication. Benchmarks covered more than 100 problem instances. On most large instances, it ran between 1.2x and 2.5x faster than the competing D-PDLP implementation.

The gains depend on problem structure and size: Nvidia's solver was slower than D-PDLP on the three ultra-large benchmark instances.

## Sources

- [NVIDIA Technical Blog](https://developer.nvidia.com/blog/scaling-decision-optimization-to-100-million-variables-and-beyond-with-mpdlp-in-nvidia-cuopt/)

---
Canonical: https://techandbusiness.org/newswire/K-rNFl6QIvj1zIFSCgBrbE
Published: 2026-10-08T02:35:19.319Z
Story chronology: 2026-10-07T15:45:00.000Z
Retrieved: 2026-10-08T04:42:53.790Z
Publisher: Tech & Business (techandbusiness.org)