AMD rocOpt brings Nvidia cuOpt-style GPU solvers to Instinct, minus one key method

AMD has introduced rocOpt, an open-source engine that runs decision-optimization workloads on Instinct GPUs. According to the official ROCm blog post, it is ready for early users on MI300X and MI355X with ROCm 7.2.3.

Four solver families on one ROCm stack

rocOpt targets linear programming (LP), mixed-integer linear programming (MILP), quadratic programming (QP) and vehicle routing. AMD says it needs no proprietary license and no platform switch for teams that already run AI and analytics on Instinct hardware.

The post lists four solver families:

• PDLP, a first-order GPU method for very large, sparse LPs. AMD calls it the recommended choice for most large-scale problems.
• Dual simplex, which runs on the CPU, suits small and medium LPs and serves as the relaxation engine inside the MILP search.
• MILP, a hybrid design. Heuristics such as local search, feasibility pump and feasibility jump run on the GPU to improve the primal bound, while branch-and-bound runs on the CPU for the dual bound. The two halves share integer-feasible solutions.
• Vehicle routing (TSP, VRP, PDP), a metaheuristic that builds an initial population of routes and improves them within a time budget set by the user.

Underneath, AMD says rocOpt builds on its Data Science stack: HIP, rocThrust, hipCUB, rocPRIM, hipRAFT, hipMM, hipSPARSE, hipBLAS and hipSOLVER. Source code, examples and build instructions are in the rocOpt GitHub repository, prebuilt containers are on Docker Hub, and a compatibility guide lists supported platforms and ROCm versions. The examples cover LP, MILP and routing through Python, CLI and Server APIs.

AMD pitches it at last-mile delivery planning, power-plant dispatch, portfolio rebalancing and supply-chain network design. The claim is that GPU solves can shorten loops that now take minutes, though the post offers no figures to back that.

A drop-in path for cuOpt users

The most practical detail sits in the resources section. AMD says rocOpt keeps API compatibility with the upstream cuOpt project, so existing cuOpt workloads can move over with minimal code changes. That makes the target audience plain: teams that already wrote against Nvidia’s engine and want to test Instinct hardware without a rewrite.

No barrier solver and no benchmarks yet

The announcement leaves several gaps, and the largest is the barrier (interior-point) LP solver. AMD says it is unavailable because the upstream implementation depends on cuDSS, Nvidia’s sparse direct solver library, and calls broader sparse direct-solver support an opportunity for future ROCm work. Users porting cuOpt code are told to switch to PDLP or dual simplex.

That matters because Nvidia has leaned on the barrier method. According to Nvidia’s developer blog, its internal benchmarks show cuOpt barrier averaging more than 8x faster than a leading open-source CPU solver and more than 2x faster than a popular commercial CPU solver on a public set of large LPs. Those are vendor-run numbers, but they show what rocOpt users give up on LPs where barrier would win. Nvidia’s launch material also cited an average 70x gain for its PDLP over a CPU-based PDLP solver, which is the method rocOpt does offer, but AMD publishes no comparable measurement for its port.

The hardware scope is narrow too. rocOpt is listed for two Instinct accelerators and one ROCm release. By comparison, cuOpt’s repository lists support for CUDA 12 and 13 and any Nvidia GPU from the Volta generation (compute capability 7.0) onward, under the Apache-2.0 license. The rocOpt post does not state its own license, and the GitHub link points to an amd-integration branch of a repository under the AMD-Ecosystem organization, which suggests the work is still being merged and is not yet a standalone release.

The MILP solver also depends on the CPU for branch-and-bound, so scaling depends partly on the host processor. Whether the hybrid split performs well on MI355X systems is something the post does not show. AMD has been selling its own CPU story for agent workloads, as in its Helion open rack blueprint, but it does not connect the two here.

The source also does not establish solve times, problem sizes handled, memory requirements or any comparison with Nvidia hardware. Until AMD or independent users publish such numbers, rocOpt is best read as a compatibility and ecosystem move: it lets the cuOpt interface run on Instinct GPUs, with the performance case still to be made.