See which processes run 56% faster on NVIDIA Vera to understand how to balance performance across the system.
When choosing hardware for Ansys LS-DYNA nonlinear dynamics structural simulation software, counting processor cores is an easy place to start, but often not a very useful place to finish. What matters is how quickly the complete system runs the simulations we care about. A processor with more cores can still be slower if those cores spend too much time waiting for data or communicating with each other. Processor speed, memory bandwidth, memory latency, cache, communication overhead, and the amount of work assigned to each core all contribute to the performance.
Recent testing of explicit LS-DYNA workloads on the new NVIDIA Vera CPU provides an interesting example. The results reinforce something we have seen repeatedly when evaluating hardware for engineering simulation: balance across the system matters more than any individual specification.
Finding the Right Balance
Explicit LS-DYNA workloads such as crash, drop test, and impact typically execute a very large number of small time steps. High sustained processor performance is important, but the processor also needs data quickly enough to keep its cores productive. Memory bandwidth and cache therefore play important roles, and adding more cores without sufficient resources can eventually produce diminishing returns.
There is also a communication cost as simulations scale. LS-DYNA software can distribute a model across many processes using the Message Passing Interface (MPI). Each additional process receives a smaller portion of the model, but the processes also need to exchange information. For explicit LS-DYNA simulations, we typically recommend distributing at least 10,000 to 20,000 elements per MPI process as a useful rule of thumb for efficient scaling.
This provides useful context for Vera. Rather than simply asking how many cores a processor has, the more relevant question is whether those cores receive enough compute performance, cache, and memory bandwidth to remain productive as LS-DYNA software scales.
Putting NVIDIA Vera to the Test
Each NVIDIA Vera processor in our test system has 88 NVIDIA Olympus cores. The system provides 1,200 GB/s of peak memory bandwidth per socket, translating to 3x greater memory bandwidth per core compared to competitive CPUs.
LS-DYNA 16.3 was tested with the same code revision for the Arm and x86 builds. The Arm build was compiled specifically for the Olympus CPU. OpenMPI 4.0.X and single precision were used, all threads were pinned, and processes in undersubscribed runs were distributed evenly across the sockets.
Using the ODB-10M explicit benchmark for 30,000 cycles, dual-socket Vera completed the simulation in 1,869 seconds using 176 physical cores. The full dual-socket competitive system completed it in 2,029 seconds using 256 physical cores. Despite using 80 fewer cores, Vera completed the benchmark about 8% sooner, corresponding to the 8.6% performance advantage reported in the benchmark results. At 192 cores, the competitive system required 2,149 seconds, giving Vera a 15% performance advantage.
Understanding Where the Difference Comes From
The overall execution time tells only part of the story. For element processing, the competitive processor performs extremely well and is actually faster than Vera at several comparable process counts. Vera does not simply execute every part of LS-DYNA software faster.
The picture changes when we look at contact processing. Vera required 323.83 seconds, compared with 506.04 seconds at 192 cores and 486.6 seconds using the full 256-core competitive system. The benchmark calculates Vera's contact-processing advantage as approximately 56% and 50%, respectively.
This result is particularly relevant for crash and impact applications, where contact can represent an important part of the LS-DYNA workload as structures deform, interact, slide, and separate. A good real-world example is NASCAR, which uses LS-DYNA software to optimize the placement of SAFER barrier modules and reconstruct actual on-track crashes. These simulations help engineers investigate vehicle accelerations and potential driver injuries and use those insights to reduce risk on the racetrack.
The NASCAR example illustrates why performance in contact-intensive portions of an explicit simulation matter. Faster time to solution can provide more opportunities to investigate designs, configurations, and scenarios within the available time. The NASCAR models have not been benchmarked on Vera, so this example should not be interpreted as a Vera performance result. It simply illustrates why efficient contact processing matters in practical explicit LS-DYNA applications. Read the story about NASCAR and SAFER barriers.
Why More Cores Aren’t Always Better
The results reinforce another important LS-DYNA lesson. Using every available processor core doesn't automatically deliver the best time to solution.
Vera's full node contains 176 physical cores, compared with 256 in the competitive system. Yet the 176-core Vera system completes this ODB-10M test faster than the 256-core system.
This is also why undersubscription can be worth evaluating. Running fewer MPI processes can make more memory bandwidth and cache available to each process while reducing communication overhead. The optimum configuration depends on the workload, so representative models should be tested at multiple process counts rather than simply assuming that all available cores should always be used.
This can also have implications beyond raw performance. For engineering organizations, the objective is not to maximize the number of processor cores in use. It is to complete the required simulation workload within available engineering time, infrastructure, power, and software budgets. Performance per licensed core and energy per completed simulation can therefore be as relevant as performance per server.
What Should LS-DYNA Users Take Away?
One benchmark should never determine a hardware purchase. Explicit LS-DYNA models vary considerably in contact intensity, element formulations, communication patterns, memory access, input/output, and parallel scalability.
The most reliable approach remains to benchmark representative production models at several process counts and measure complete time to solution. Look beyond peak core count. How well does performance scale? Where does the simulation spend its time? Does reducing the number of processes increase efficiency? How much energy does the completed simulation consume? And what is the infrastructure and software cost of getting that engineering result?
The new LS-DYNA 16.3 results provide an interesting data point. One hundred seventy-six Vera physical cores completed ODB-10M faster than 256 cores in the current competitive system, despite the latter being faster in element processing. Vera's much stronger contact-processing performance more than compensates for that difference in this workload.
That is perhaps the most useful lesson from these results. Don't choose LS-DYNA hardware simply by counting cores or comparing one processor specification. Measure how effectively the complete system runs the simulations that matter to you.
Interested in what these results could mean for your LS-DYNA workloads? Visit us at Booth #1727 at Supercomputing 26 and meet with our LS-DYNA experts. We would be pleased to discuss the Vera results, your own explicit simulation workloads, and what to consider when selecting and configuring hardware for LS-DYNA.












