Full-Vehicle NVH: The Bottleneck Is the Clock—GPU Acceleration in MSC Nastran

How MSC Nastran FastFR and GPU acceleration take multi-million DOF NVH runs from overnight to the same day.

Full-vehicle NVH models keep getting bigger, and electric programs push the analysis higher in frequency at the same time. A modern body-in-white can run to tens of millions of degrees of freedom, and the frequencies that matter now reach well past where combustion masking used to hide them. Run a wide-band frequency response on a model like that, and the limiting factor is rarely the physics. It is the clock. When a single production run ties up a machine overnight, the number of design iterations an NVH team can fit into a program is capped by runtime rather than by engineering ideas.

Figure 1: Cabin sound quality is the endpoint every full-vehicle NVH study works back from.

Why Frequency Response Becomes the Bottleneck

Modal frequency response works in two stages. First, an eigensolution extracts the modes; in MSC Nastran, this is the job of ACMS, the workhorse eigensolver for large NVH models. Then the solver computes the response across the frequency sweep by working in that modal space. The second stage is where large models get expensive. As the model grows and the band extends upward, both the mode count and the number of frequency points climb together. On the benchmark below, that meant 50,000 modes evaluated across 1,500 frequencies. The dense matrix operations that combine all of that, dominated by matrix multiply, scale hard with both numbers, and they come to own the wall-clock time.

Where GPUs Change the Economics

FastFR is the accelerated frequency-response path in MSC Nastran, and it pairs directly with ACMS. It has been GPU-enabled, and the reason it maps well onto GPU hardware is specific: the dense linear algebra that dominates the runtime is exactly the kind of operation GPUs are built to accelerate. Rather than rewriting the workflow, the GPU offloads the heaviest numerical work while the rest of the solve proceeds as normal. GPU-accelerated FastFR scales across multiple GPUs, so a machine with two cards can push further than one.

The Benchmark

The reference benchmark for this comes from results presented in "A Deep Dive into NVH Analysis Using MSC Nastran": a 43.5 million-degree-of-freedom automotive body model, solving 50,000 eigenmodes across 1,500 frequencies. With GPU-accelerated FastFR, the configuration showed up to 50 percent faster runtime on the runs that extended to high frequency. Broken down by operation, the dominant matrix-multiply step showed roughly a 15X improvement, and the FastFR stage overall came down by about a factor of two on the reported configurations, with further gains moving from one GPU to two.

The pattern inside those numbers is the part worth carrying into a program decision. The benefit grows with the frequency range. The higher you push the analysis, the more modes and frequency points enter the calculation, and the more of the runtime lands on the operations the GPU accelerates. The runs that used to be the most painful are the ones that speed up the most.

Figure 2: GPU-accelerated FastFR in MSC Nastran on Volvo Car Corporation's 43.5 million DOF body model, solving 50,000 modes across 1,500 frequencies.

Why this Matters Most for EV Programs

This is not an incremental convenience for electric vehicle work; it is aimed at the exact regime EVs create. Motor whine and inverter harmonics reach into the kHz range, which means more modes and more frequency points in every full-vehicle run, precisely the conditions under which GPU acceleration returns the most. A team that has moved its NVH targets upward in frequency to chase e-drive tonal noise is running the workload that benefits from this the most, not the least.

Figure 3: Electric drive units concentrate NVH energy where it is hardest to compute. High-frequency electromagnetic forces excite the housing up to 11 kHz, and the surrounding coolant couples strongly with the structure.

The business case follows from the runtime. Cutting a high-frequency full-vehicle run in half means either the same study is finished in half the queue time, or twice as many design variants are evaluated in the same schedule. On programs where NVH signoff sits on the critical path, that is iterations recovered rather than compute saved for its own sake, and it changes what GPU hardware is worth. The capex pays back through throughput on the models that were previously gating the schedule.

Getting It Running

Because FastFR pairs with ACMS, and ACMS is built into MSC Nastran, the entry point is the solver most NVH teams already run. The path from a CPU frequency response workflow to a GPU-accelerated one is a configuration change on hardware you provision, not a re-authoring of the model. That low switching cost is part of why the Volvo result is a useful reference point rather than a special case.

If you want to size the gain for your own models, the shape of your workload tells you most of the answer up front. The larger the model and the higher the frequency band, the more there is to recover.

For more on the diagnostics, coupled acoustics, and HPC methods behind modern NVH work, our "Engineer's Guide to Modern NVH Simulation" walks through the full workflow:

You can also subscribe to get new technical guides as they are published.