Manufacturing

Physical testing is slow, expensive, and late-cycle. By the time a prototype reveals a structural weakness, an overheating problem, or a casting defect, the development budget has already been spent and the schedule has slipped. Manufacturers that shift validation upstream — into simulation — compress development cycles, reduce physical test iterations, and catch failure modes before tooling is committed.

The constraint is compute. A detailed FEM crash analysis, a transient thermal simulation of an industrial assembly, or a coupled mold-filling and solidification calculation for a complex casting all demand more core count, memory, and I/O bandwidth than a workstation cluster can reliably provide. HPC infrastructure removes that constraint and makes simulation-first development viable across the product program.

Key Workloads

Manufacturing simulation spans structural, thermal, fluid, and process domains. Each has distinct software and hardware requirements.

Application Areas

Structural Strength and Durability Analysis

Structural simulation covers a wide range from simple static loading through nonlinear contact, large deformation, impact, and long-cycle fatigue. Large assemblies — gearboxes, frames, pressure vessels, suspension components — routinely produce FEM models with 5–20 million elements. These models require both high core counts for solve speed and high per-node memory to hold the global stiffness matrix in RAM.

Static and nonlinear analysis (Abaqus, ANSYS Mechanical): Typical production runs use 32–128 cores. Nonlinear cases with contact and plasticity are memory-bound; 512 GB per node is a practical minimum for large assemblies.

Explicit dynamics (LS-DYNA, Radioss): Drop tests, impact, and sheet metal forming use explicit time integration. These scale nearly linearly with core count and benefit strongly from InfiniBand — inter-node communication is frequent and latency-sensitive.

Fatigue analysis: fe-safe and nCode DesignLife post-process FEM stress results to predict service life under cyclic loading. Computational load is moderate but high storage throughput is needed to stream large result databases.

Thermal Management and Heat Transfer

Industrial equipment — motors, drives, power electronics, compressors, heat exchangers — requires careful thermal validation. Overtemperature failures are field failures; simulation-driven thermal design prevents them.

Thermal simulation workloads include:

• Steady-state conjugate heat transfer (CHT): Cooling fin geometry optimization, forced-air cooling duct design
• Transient thermal analysis: Cyclic loading effects, startup/shutdown thermal gradients
• Electronics cooling: PCB-level and enclosure-level thermal management (ANSYS Icepak, COMSOL)
• Coupled electromagnetic-thermal: Motor and transformer simulation (ANSYS Maxwell + Mechanical)

COMSOL Multiphysics coupled simulations are particularly memory-intensive. Nodes with 256–512 GB RAM and fast local NVMe scratch are the recommended hardware profile.

Fluid Mechanics: Pumps, Valves, and Pipe Systems

Internal flow simulation for hydraulic and pneumatic components — pumps, control valves, pressure regulators, heat exchangers, and piping networks — is a core CFD application in machinery manufacturing.

Key workflows:

• Pump characteristic curve generation: CFD replaces or supplements physical performance testing; rotating machinery simulation using MRF or sliding mesh
• Valve flow coefficient (Cv) prediction: Pressure drop and cavitation analysis across operating range
• Pipe system pressure loss: Network-level flow distribution; multiphase and cavitation effects
• Compressor internals: Stage-by-stage aerodynamic analysis (ANSYS CFX, OpenFOAM)

These simulations scale well with core count and require InfiniBand for efficient MPI communication across nodes. OpenFOAM-based workflows benefit from 32–256 cores per run; commercial solvers (Fluent, CFX) perform well at similar scales.

Mold Design and Casting Simulation

Casting defect prevention — porosity, shrinkage, cold shuts, hot tears — requires simulation of both mold filling and solidification. Getting this wrong means scrapped castings or, worse, field failures in structural parts.

MAGMASOFT and ProCAST are the industry-standard tools for foundry simulation. Both are CPU-intensive with high memory requirements, and ProCAST benefits from InfiniBand for large-domain parallel runs.

Typical casting simulation workflow:

1. Mold filling: Track melt front, air entrapment, temperature distribution during fill
2. Solidification: Predict shrinkage porosity location and hot spot formation
3. Thermal stress: Residual stress and distortion in the final casting
4. Gating and riser optimization: Iterate on feeding system design to eliminate defect zones

A full casting simulation for a complex part (engine block, differential housing, structural bracket) can require 64–256 CPU cores and 6–24 hours of wall time.

Digital Twin Applications

Digital twins couple real-time sensor data from operating equipment with physics-based or data-driven simulation models to predict remaining useful life, detect anomalies, and optimize maintenance intervals.

HPC is involved at the model development stage: full-fidelity FEM and CFD simulations generate the training data or reference solutions that reduced-order models (ROMs) are built from. GPU-accelerated machine learning (PyTorch, TensorFlow) trains the ROM once high-fidelity simulation data is available.

Mevasis provides the HPC infrastructure for both the high-fidelity simulation phase and the GPU training phase of digital twin development.

Typical HPC Configuration

Reference configuration for a manufacturer running structural, thermal, and casting simulations:

Login / Pre-/Post-Processing Nodes (2×)
├── CPU Compute Nodes (8–32 units)
│   └── 2× AMD EPYC 9654 (96 cores/node), 512 GB DDR5
│       (Abaqus, LS-DYNA, Fluent, OpenFOAM, ProCAST)
├── High-Memory Nodes (2–4 units)
│   └── 2× EPYC, 1–2 TB DDR5
│       (large FEM assemblies, COMSOL coupled, MAGMASOFT)
├── GPU Nodes (optional, 2–4 units)
│   └── 2× Intel Xeon + 4× NVIDIA L40S
│       (ANSYS GPU solver, ROM training, post-processing)
└── Scratch Storage
    └── BeeGFS or Lustre, NVMe + SAS hybrid
        (10–50 TB fast scratch, 200+ TB archive)

Scheduler: SLURM with per-software partition configuration
Network: InfiniBand HDR100 for CPU cluster backbone
License server: Dedicated FlexLM / DSLS host, HA pair recommended

For smaller teams or project-based needs, HPC Rental provides access to the same hardware profile without capital expenditure.

Mevasis Manufacturing HPC Services

Mevasis has direct experience sizing and deploying HPC clusters for FEM, CFD, and casting simulation workloads. The team understands the license architecture of commercial simulation software and can configure SLURM to correctly track token consumption for ANSYS, Abaqus, and LS-DYNA.

• Workload profiling: Benchmark your existing simulation jobs to determine optimal cluster sizing
• Turnkey installation: Hardware procurement, InfiniBand network design, SLURM and license server configuration
• Commercial software support: ANSYS HPC Pack configuration, Abaqus token optimization, LS-DYNA MPI tuning
• HPC Rental: Project-duration or annual cluster rental — no capital commitment required
• HPC Consulting: Performance audits, bottleneck identification, storage architecture for large result datasets

Frequently Asked Questions

How many cores should we use for large FEM simulations? It depends on solver type. Linear static analysis (ANSYS Mechanical, Nastran) scales well to 32–64 cores; beyond that, communication overhead often limits returns unless InfiniBand is present. Explicit dynamics (LS-DYNA) scales much further — 128–512 cores with near-linear efficiency on InfiniBand. Nonlinear implicit runs are memory-bound; adding more cores helps only if memory per node is adequate.

We use ANSYS and Abaqus. Will our existing licenses work on an HPC cluster? Yes, provided your license type includes HPC parallel execution rights. ANSYS uses HPC Packs or token-based licensing; Abaqus uses token licensing. Mevasis performs license audits as part of cluster scoping to ensure you are not purchasing hardware that your license terms prevent you from using fully.

Is on-premise or rental better for our manufacturing team? If simulation is continuous (daily usage across multiple engineers), on-premise ownership is typically more economical over a 3–5 year horizon. If simulation is project-driven — intensive during new product development, light otherwise — rental is more cost-efficient and avoids idle capital. Mevasis provides both options and will help you assess which fits your utilization pattern.

Can casting simulation software run efficiently in parallel? ProCAST parallelizes well via MPI for large casting domains; InfiniBand improves efficiency at 32+ cores. MAGMASOFT has a more limited parallelization model and is better suited to high-memory single-node or small-SMP configurations. Mevasis configures SLURM partitions appropriately for each tool.