What is rocOpt?

What is rocOpt?#

2026-10-13

2 min read time

Applies to Linux

rocOpt is an open-source, GPU-accelerated engine for decision optimization on AMD Instinct GPUs through the ROCm software stack. It solves large-scale linear programming (LP), mixed-integer linear programming (MILP), quadratic programming (QP), and vehicle routing problems containing millions of variables and constraints, returning near real-time results on AMD Instinct MI300X and MI355X GPUs.

rocOpt is aligned with and API-compatible with cuOpt 26.02, so you can run existing cuOpt workloads and optimization pipelines on AMD Instinct GPUs without rewriting client code.

By moving the heavy linear algebra onto the GPU, rocOpt accelerates decision-making in operations research, logistics, supply chain, and scheduling workloads — domains where traditional CPU-only solvers often struggle to deliver answers within useful time budgets.

The current release is rocOpt 1.0.0, which is compiled against ROCm 7.1.1 and runs on the ROCm 7.2.3 runtime. This includes the following features:

  • rocOpt offers solver families tuned for the main mathematical programming problem classes. Each solver targets a different balance of problem size, structure, and solution quality:

    • PDLP (Primal-Dual Hybrid Gradient) — GPU first-order method for very large, sparse LPs; recommended for million-variable problems on rocOpt.

    • Dual simplex — CPU-based method for small-to-medium LPs; also drives LP relaxations inside the MILP branch-and-bound loop.

    • MILP — hybrid GPU/CPU mixed-integer solver combining GPU primal heuristics with CPU branch-and-bound.

    • QP — quadratic programming for objectives such as portfolio risk/return trade-offs.

    • Vehicle routing (TSP / VRP / PDP) — metaheuristic routing that improves routes within a user-specified time budget.

  • rocOpt includes APIs and tooling to build, run, and deploy optimization workflows:

    • Python API — construct models, configure solvers, and run solves from Python (import cuopt for API compatibility with upstream).

    • CLI — solve problems from standard MPS/LP inputs on the command line.

    • Server API — serve optimization requests over HTTP with a cuOpt-compatible interface.

  • rocOpt is built on the AMD Data Science stack — a HIP port cuOpt that maps RAPIDS-style dependencies to their ROCm equivalents (hipRAFT, hipMM, rocThrust, hipCUB, rocPRIM, hipSPARSE, hipBLAS, hipSOLVER) so workloads run on AMD Instinct hardware alongside other AMD Data Science libraries such as hipDF.

  • rocOpt supports concrete decision-optimization use cases where GPU acceleration changes what is tractable in production, including:

    • Last-mile logistics and fleet routing (VRP) with time windows and capacity constraints.

    • Energy dispatch and unit commitment (MILP) with short market intervals.

    • Portfolio optimization (QP) over thousands of instruments.

    • Supply chain network design (MILP) across multi-echelon networks.

Note

The Barrier (interior-point) LP solver depends on the cuDSS sparse direct solver, and has no ROCm equivalent in this release. Use PDLP for large LPs and dual simplex for smaller problems.