GPU-native explicit finite-element solver

More physics.
Less waiting.

Put GPU computation to work on your next engineering question.

MinuteSim brings GPU-resident explicit simulation to medium-to-large structural workloads. Explore more designs. Build toward richer physics data for AI.

Explore simulations Discuss an evaluation

Actual S-rail forming result, colored by effective plastic strain at the shell top fibre, with the original 0 to 0.40 color scale. — ACTUAL MinuteSim RESULTA sheet becomes a structure.
EXPERIMENTAL
Adaptive shell refinement

GPU-residentExplicit time integration

Shells · solids · beamsSupported configurations

FP32 & FP64Single and double precision builds

01 / See the physics

Real models.
Actual simulation results.

Forming, impact, bending and large deformation. Explore results generated by MinuteSim, with their original physical fields and color scales.

01 / 08

S-rail forming

Follow a sheet as it is drawn into an S-shaped rail. Color shows effective plastic strain at the shell top fibre.

Result field: Effective plastic strain
Evidence class: EXPERIMENTAL · adaptive refinement

Capability demonstration. The animation shows the evolving mesh; the hero image shows the result without mesh lines. This is separate from the runtime case below.

Open video

These clips demonstrate configured capabilities. Independent numerical validation is documented separately. See validation scope ↗

02 / The value of shorter runtimes

Large models.
A different timescale.

Make room for the next design iteration. This solid-compression scenario shows how the CPU comparison basis changes the runtime ratio.

EXPERIMENTAL · ILLUSTRATIVE SCENARIO

590×

Illustrative CPU-to-GPU runtime ratio

Supplied timing assumptions · one configuration

SOLVER B / 1 CPU THREAD22.6 hoursEstimated from a short timing window

MinuteSim / ONE NVIDIA L402.3 minutes137.8 s projected at normal GPU clocks

1,886,592 solid elements0.5 s physical analysis duration119,821 MinuteSim time increments*

FP32 GPU / extended-single CPU configuration on separate EPYC 9274F-class servers. Time increments and output differ. Final-field equivalence is not established for this timing case. *Increment count comes from the corresponding completed GPU run.

Read the assumptions and observed runtimes

A scenario for one model, not a product-wide speed guarantee. The observed 8-thread CPU and GPU full-run times are available alongside the projections.

EXPERIMENTAL · ILLUSTRATIVE SCENARIOS

Runtime scenarios across model sizes.

One NVIDIA L40 versus Solver B. Compare the same problem families with 8 CPU threads and 1 CPU thread.

Deformed sheet over a rounded punch, colored by shell-top effective plastic strain, with the original 0 to 0.25 color scale. — Sheet forming
A sheet stretches over a rounded punch.

Four bent specimens made with triangle shells, quad shells, tetrahedra and hexahedra, with tooling and the original effective-plastic-strain color scale. — Mixed-element bending
Shell and solid specimens bend under tooling.

Quarter view of a rigid hemisphere compressing a solid block, colored by effective plastic strain with the original 0 to 0.50 scale. — Solid compression
A rigid hemisphere presses into a solid block.

Match each model title’s color to its curve below. Images show representative results; the curves cover multiple model sizes.

8 CPU thread runtime scenarios. At the largest plotted sizes, CPU-to-GPU ratios are 132 times for sheet forming, 46 times for mixed-element bending and 92.4 times for solid compression. Both axes are logarithmic. — 8 CPU threads / one L40View full size ↗

1 CPU thread runtime scenarios. At the largest plotted sizes, CPU-to-GPU ratios are 517 times for sheet forming, 314 times for mixed-element bending and 590 times for solid compression. Both axes are logarithmic. — 1 CPU thread / one L40View full size ↗

X: element count. Y: CPU time / GPU time. Both panels use the same logarithmic axes.

Illustrative scenarios using supplied timing assumptions, including recorded timings and estimates. Endpoint labels refer to the largest plotted model in each family. Bending timing uses prescribed motion without contact; the gallery shows separate contact examples. Comparison conditions and data ↗

03 / Physics data for AI

1,000 simulations
for AI training.

Representative physics data is part of the foundation for surrogate models. Faster simulation can change the cost of exploring that data space.

PLANNED APPLICATION

Explore the compute-time scenario.

Multiply a recorded single-run time by a chosen number of sequential simulations.

Simulations for training data 1,000

1002,000

MeshGraphNets used 1,000 training trajectories per dataset. It is an example of dataset scale, not a universal requirement. Read the study ↗

Sheet forming

19,881 shells

CPU · 8 threads

79 days

One L40 GPU

18.5 hours

Per run: CPU 6,839.55 s · GPU 66.676 s

Mixed-element bending

64,800 elements · contact

CPU · 8 threads

16.5 days

One L40 GPU

2.6 days

Per run: CPU 1,429.374 s · GPU 226.817 s

0 days50 days100 days

Both charts share a linear scale. CPU: Solver B on 8 EPYC 9274F threads. GPU: one NVIDIA L40. Single-precision configurations.

Illustrative arithmetic, not measured sustained throughput or a completed dataset. Sequential runs on one CPU allocation or one GPU; concurrent CPU jobs would change the comparison. Data preparation and AI model training time are excluded. Scenario details ↗

The hardware opportunity

GPU hardware.
A wider horizon for physics data.

Selected hardware generations show how compute and memory bandwidth have developed. This creates room to explore larger models and more physics data for AI.

GPUCPUL40 / benchmark GPU

FP32 compute for selected desktop CPUs and GPUs from 2008 to 2025, on an absolute TFLOPS axis. Endpoints: RTX 5090 at 104.9 TFLOPS and Ryzen 9 9950X at 4.4 TFLOPS. The separate L40 reference is 90.5 TFLOPS. — FP32 compute / selected desktop productsView full size ↗

Memory bandwidth for selected data-center GPUs and server CPU sockets, 2007 to 2025, on an absolute GB per second axis. Endpoints: B300 at up to 8,000 GB/s and Xeon 6980P at 844.8 GB/s. The L40 reference is 864 GB/s. — Memory bandwidth / GPU or CPU socketView full size ↗

The MinuteSim comparisons on this page use NVIDIA L40.

SUPPORTED hardware specifications. Absolute linear axes. GPU values follow published FP32 conventions; CPU compute values are theoretical at base clocks. Growth labels use each series’ own starting value. B300 bandwidth is up to 8,000 GB/s for the selected configuration. The L40 marker uses 2023 OVX system availability. These specifications do not predict solver speed or compare matched price and power. Products, conventions and sources ↗

Historical source data: Karl Rupp, CPU, GPU and MIC Hardware Characteristics over Time, CC BY 4.0, and Epoch AI, Data on machine learning hardware (CC BY 4.0), plus selected vendor specifications. Figures redrawn with absolute axes and an added L40 reference; graph extracts from the v11 presentation.

04 / The solver today