MinuteSim
← Performance

Performance / CPU scaling scenario

One GPU.
How many CPU servers?

Translate the runtime advantage into an equivalent CPU server count, using a trend fitted to published Solver A scaling results.

01 / One NVIDIA RTX 5090

CPU server equivalents
across model sizes.

Each server represents 128 active CPU cores. The three model families connect the runtime comparison to a CPU scaling scenario.

Sheet forming

16 servers

50,176 shells · Nakajima

Trend band: 13–20 servers

Solid compression

28 servers

1,886,592 elements · long-stroke dome

Trend band: 23–34 servers

Mixed-element bending

11 servers

494,080 elements · four families

Trend band: 9–14 servers

Swipe to explore the chart →

Derived CPU server equivalents by model size for sheet forming, solid compression and mixed-element bending. Largest plotted models: 16, 28 and 11 servers.
128 CPU cores per server · point labels round up to whole serversFull-size chart ↗

Runtime ratios use Solver B on one AMD EPYC 9274F CPU thread and MinuteSim on one RTX 5090. Explore the runtime comparisons →

02 / Published CPU scaling

How CPU performance
scales with core count.

Five published series, 33 data points. A fitted trend connects core count with speedup across the selected studies.

Swipe to explore the chart →

Five selected published Solver A scaling series, a fitted power-law trend, residual band and ideal linear scaling. Series without a one-core run use fitted starting levels.
Selected published results · fitted trend and residual bandFull-size chart ↗

Speedup ≈ 1.03 × N0.845
N = CPU cores / MPI processes

The fitted trend gives about 112× at N = 256 and 200× at N = 512. The dotted line shows ideal one-for-one scaling.

03 / Calculation method

From runtime ratio
to server count.

1

Compare runtimes

Divide the Solver B single-thread runtime by the RTX 5090 runtime for the same model.

2

Apply the CPU trend

Find the core count that gives the same speedup on the fitted Solver A curve.

3

Convert to servers

Divide the core count by 128 and round up to a whole number of servers.

Read the fit, source conventions and scenario assumptions

The fit minimizes squared residuals in log speedup. Two series start at one core. The other three start at 4, 32 and 64 workers; their starting levels are fitted jointly with the exponent and coefficient. Their plotted one-core speedups therefore include a fitted normalization.

The shaded band is the fitted trend multiplied and divided by about 1.19 (twice the residual standard deviation in log space, using 28 residual degrees of freedom). It contains 32 of the 33 points. This describes scatter around the selected fit rather than a prediction interval for a particular cluster.

The server scenario applies the Solver A trend to Solver B runtime ratios. The selected series reach 512 CPU cores / MPI processes; larger server counts extend that trend. Model size per worker, processor generation, solver configuration and communication performance differ between the source studies and the MinuteSim comparisons.

For the conversion, one MPI process represents one active CPU core. One server is a counting convention of 128 active CPU cores (two 64-core processors). The curve supplies the scaling assumption; no specific 128-core server was timed for these equivalents. The underlying CPU single-thread values include recorded times and short-window full-run estimates. Runtime methods and conditions →

This estimate uses the supplied GPU times before rounding. The Performance charts display times to one decimal place; using those rounded values gives the same whole-server counts.

N = (runtime ratio / a)1/b; servers = ceil(N / 128), using the full-precision fitted a and b.

04 / Sources

Published scaling results.

The charts are replotted from the selected numerical series. The links below lead to the original publications.

  1. Y.-S. Roh (2000). Scaling Study of LS-DYNA MPP on High Performance Servers ↗

    Neon · Sun E10000 · selected default-decomposition series.

  2. Y.-Y. Lin and C.-D. Kan (2000). A Performance Study of LS-DYNA on Vehicle Crash Simulation ↗

    Neon · HP V2500 · 1.7 million degrees of freedom.

  3. J. O. Hallquist (2006). Recent Developments in LS-DYNA ↗

    Three-car model · 750 K nodes · Cray XD1.

  4. J. Wang (2022). LS-DYNA High Performance Computing ↗

    ODB-10M · AWS EPYC · selected TCP-fabric results without EFA.

  5. Dell Technologies (2023). Scaling Ansys LS-DYNA with Cornelis Omni-Path Express ↗

    car2car · EPYC 7713 · HDR InfiniBand series · 64 active MPI processes per 128-core node.

See the time behind the comparison.

Runtime and speedup for each model size, plus GPU-to-GPU comparisons.