📘 Resource Overview & Documentation Details

You are viewing the technical directory listing for HPC for Digital Manufacturing: Altair Performance. This engineering resource is curated under our Digital manufacturing division to support automation specialists, control systems engineers, and manufacturing technicians looking for verified reference materials.

🔍 Contextual Abstract & Technical Scope

Share Summary:LinkedInEmail

High-performance computing (HPC) for digital manufacturing can help engineering teams run computationally demanding simulations for virtual product development. But cluster performance depends on more than processor core count: the solver, model size, memory and storage needs, network, and parallel configuration all matter. Benchmark results for Altair AcuSolve and Altair Radioss on Dell EMC systems with AMD EPYC processors illustrate why workload-specific testing is important when designing a CAE cluster.

HPC digital manufacturing starts with workload fit

A building-block approach allows an HPC system to be assembled around the requirements of its users and applications. Compute nodes provide the resources to run simulations, while infrastructure servers handle system administration and user access. Storage and networking complete the environment, and cluster management and job scheduling software support deployment and workload operations.

This modular design is useful because manufacturing workloads are not uniform. A cluster intended for a particular mix of solvers and simulation sizes may need a different balance of compute, memory, local storage, and interconnect than one built for other jobs. The configurations and benchmark results described here relate to tests conducted in 2019, using specific hardware and software versions; they should be treated as workload-specific evidence, not as a guarantee of performance for other environments.

Designing compute, storage, and network building blocks

Compute building blocks run CAE simulations, and the processor and memory configuration should reflect the expected application mix. The tested cluster included PowerEdge C6525 servers with dual AMD EPYC 7452 processors, 256 GB of memory, and HDR100 InfiniBand. Additional single-node tests used systems with EPYC 7402, 7502, and 7702 processors. This provided a basis for comparing processor configurations as well as scaling jobs across nodes.

Storage design also affects the user experience. Scratch storage holds temporary simulation data, often for the duration of a run. For many CAE workloads, direct-attached storage on compute nodes can provide a suitable balance of cost and performance. However, workloads with greater capacity or I/O demands may require different sizing or a shared high-performance file system. Operational storage holds project data and results for ongoing access; archival storage is intended for longer-term retention. These tiers serve different purposes and should be planned accordingly.

HPC clusters typically separate system administration traffic from the high-speed fabric used for node-to-node communication. The former supports provisioning and management; the latter can be important as jobs are distributed across servers. The right network depends on cluster size and workload communication needs. Scheduling software also helps coordinate access to compute resources. In the described solution, Altair PBS Professional provides job scheduling capabilities, while Bright Cluster Manager is identified for cluster installation and monitoring.

Altair AcuSolve performance: model size changes scaling

AcuSolve is a computational fluid dynamics solver used for CFD and multiphysics applications. In the reported single-server benchmarks, the EPYC 7502 and 7452 systems performed well, while the 64-core EPYC 7702 system provided a notable advantage over the 32-core systems in the tested cases. The size of that advantage varied by model, underscoring that processor comparisons should use representative simulations rather than core counts alone.

Multi-node results showed different scaling behavior across the Riser, Windmill, and Nozzle models. Nozzle, the largest model in the set, scaled nearly linearly through eight nodes. The other models showed more limited improvement at four nodes or above. Riser also achieved an unusually strong initial gain from one to two nodes. The analysis attributed this to cache effects: distributing a problem can allow its data to fit more effectively in cache. Such effects are specific to the model and do not imply that adding nodes will always produce the same gain.

Choosing a hybrid parallel configuration

AcuSolve can combine shared-memory threading within a node with distributed-memory parallelism across nodes. In these tests, adding threads per domain offered little benefit until runs used four nodes; at eight nodes, multiple threads helped for some models. The Nozzle model showed strong speedup across the thread configurations tested. A suggested starting point was to use a number of threads corresponding to the number of nodes, while validating settings against each workload rather than treating this as a universal optimum.

Altair Radioss performance: test against the simulation

Radioss is a structural analysis solver for highly nonlinear problems under dynamic loading. In the single-server comparisons, the second-generation EPYC systems improved on the first-generation systems in the tested cases, and the EPYC 7702 delivered the strongest overall performance among the compared systems.

Scaling again depended on model size. The larger Taurus model achieved nearly linear speedup through eight nodes, while the smaller Neon model’s gains began to diminish beyond two nodes. The Neon simulation was relatively small, so distributing it over more servers offered less opportunity for efficient scaling. A larger node count is therefore not automatically a better choice for every job.

Hybrid parallel results differed from AcuSolve. Taurus performed best with one shared-memory thread in the tested configurations, while Neon showed a modest benefit from additional threads at four or more nodes. This contrast is a practical reminder that hybrid settings should be benchmarked separately for each solver and model.

Practical implications for CAE cluster planning

  • Benchmark representative jobs. Include the models and simulation sizes users actually run; results can vary substantially between cases.
  • Evaluate scaling, not just peak node performance. Measure whether a workload benefits from additional nodes and account for the point at which gains taper off.
  • Tune parallelism by application. Thread and domain settings that help one solver or model may not help another.
  • Size storage and networking to the workflow. Consider scratch I/O, shared data requirements, and communication between nodes when selecting system components.

The central lesson from these Altair benchmarks is that effective HPC for digital manufacturing is a system-design and workload-tuning exercise. AMD EPYC processor configurations delivered strong results in the tested AcuSolve and Radioss cases, but model size, cache behavior, node count, and parallel strategy shaped the outcome. Engineering teams planning a CAE cluster should use workload-specific benchmarks to guide compute, storage, network, and scheduling decisions—and validate those choices with their own simulations.

Access Official Technical PDF