Breaking the Memory Bottleneck with scFAST Solver in Cradle CFD

In large-scale computational fluid dynamics (CFD), speed is often framed around faster processors, more cores, and higher floating-point throughput. Modern CPUs and GPUs have certainly transformed what engineers can simulate within practical design timelines. But for industrial CFD models with tens or even hundreds of millions of cells, raw compute is only part of the equation. The real constraint is often memory.

Every CFD solve repeatedly accesses mesh topology, boundary conditions, matrices, and flow variables such as pressure, velocity, etc. If this data cannot be stored efficiently or moved through the hardware fast enough, performance becomes limited by data movement rather than calculation speed. Faster CFD, therefore, depends not only on compute power but also on memory capacity, bandwidth, data locality, and solver design.

This is the challenge addressed by scFAST, the new solver in Cadence Cradle CFD. Using a reduced-memory approach, the new solver supports larger simulations with a significantly smaller footprint while preserving the accuracy and convergence of the solution.

Why Memory Becomes the Bottleneck in GPU-Based CFD

GPU computing is well-suited for CFD because it offers exceptional floating-point throughput and scalability for data-parallel workloads. However, implicit unstructured-grid solvers remain difficult to scale efficiently on GPUs because they do more than execute arithmetic operations: they repeatedly access neighboring-cell information, residuals, Jacobian-related terms, solution updates, and physical variables across complex meshes.

For high-resolution polyhedral meshes, each cell can have many neighbors. In implicit methods, those couplings are represented through large, sparse Jacobian matrices containing many off-diagonal matrix blocks, which can quickly consume GPU memory and bandwidth even when the processor has enough compute capability.

This becomes a major bottleneck for large-scale compressible flow problems, where teams need added resolution around shock waves, heat-transfer gradients, tip regions, or other localized features. When available memory is exhausted, mesh refinement must be compromised. By lowering the memory needed for the solver, engineers can fit larger CFD models within the same GPU memory limit.

Introducing scFAST: A GPU-Accelerated Solver in the scFLOW Framework

Cradle CFD is a comprehensive platform for industrial applications, with scFLOW supporting automatic polyhedral mesh generation and robust solver workflows. While scFLOW has traditionally focused on CPU-based simulations using MPI parallelism, scFAST extends the environment with a GPU-accelerated capability for large-scale, memory-efficient simulation, giving users flexibility across CPU and GPU resources.

The solver is built around the reduced-memory multicolor Gauss–Seidel method, or RM-MCGS. Inspired by the block LU-SGS solver concept, it targets one of the most storage-intensive parts of conventional MCGS: the off-diagonal blocks associated with neighboring cells.

Instead of storing these off-diagonal blocks, RM-MCGS computes the required contributions on the fly using a derivative-based formulation. This reduces memory demand while retaining the same underlying numerical formulation as the conventional method. In practice, it enables larger meshes to fit within a given GPU memory budget while maintaining comparable convergence behavior.

scFAST currently supports steady, transient RANS and LES calculations, along with an MRF approach for rotating domains. Its focus is on efficient solutions for large-scale applications, including high-speed flows and selected turbomachinery use cases.

Evaluating the Accuracy of the Solver for Aerospace Applications

Memory efficiency is valuable only if accuracy is preserved. For performance evaluation of the transonic Common Research Model aircraft case, MCGS and RM-MCGS (scFAST) were compared using the HLLC flux, a less dissipative numerical flux suited to shock-wave analysis and high-speed flow applications. The results matched closely for both solvers, as shown in the figure below.

Comparing linear solver iteration count (left) and evolution of aerodynamic coefficients (right) using both MCGS and RM-MCGS solver

Compared with the simpler but more dissipative Rusanov flux, HLLC better preserves contact and shear waves, making it useful when shock resolution, pressure distribution, and wave structure matter. The key question was whether combining HLLC with RM-MCGS retains accuracy without eroding performance.

Isosurfaces of surface pressure distributions for (i) Rusanov flux (left), (ii) HLLC flux (right)

The results were encouraging; RM-MCGS reproduced the numerical behavior of the original MCGS method, while HLLC improved pressure distribution and shock-capturing accuracy when compared with Rusanov. From the figure below, it is clear that RM-MCGS reduces GPU memory usage by around 50% for both Rusanov and HLLC fluxes compared to MCGS.

A separate shock-wave diffraction example also showed improved wave resolution with HLLC. Together, these cases demonstrate that the reduced-memory approach is not only a capacity optimization; it can also support practical aerospace simulations in which shock structures and wave interactions are central to the engineering question.

Comparing the wave resolution in (i) Schlieren photograph (Bazhenova et al., 1984), (ii) Rusanov flux, (iii) HLLC flux

The method achieves approximately a 50 percent reduction in GPU memory usage. For engineering teams, this is a meaningful shift. Instead of treating GPU capacity as the limiting factor, users can now refine key regions or support more demanding physics within the same hardware envelope. This is especially relevant as CFD users move toward more detailed models and broader design exploration across aerospace, automotive, turbomachinery, and thermal applications.

What scFAST Unlocks for CFD Teams

For CFD teams, the value of scFAST is practical and immediate: it helps remove memory as a barrier to larger, more detailed simulations. By reducing the storage required for implicit solution updates, RM-MCGS enables scFAST to use GPU resources more efficiently while preserving the solver behavior engineers rely on.

As CFD models grow and engineering timelines continue to tighten, solver efficiency must be measured by how intelligently memory and compute are used together. With scFAST, Cadence Cradle CFD takes an important step toward GPU-enabled simulation workflows that can keep pace with the scale and fidelity modern engineering demands.

References

Nakashima, Y., Nishikawa, H., Lee, J., & Cerizza, D. (2025). A reduced-memory multicolor Gauss-Seidel relaxation scheme for implicit unstructured-polyhedral-grid CFD solver on GPU. AIAA 2025-3870. American Institute of Aeronautics and Astronautics. https://doi.org/10.2514/6.2025-3870

Bazhenova, T. V., et al. (1984). Unsteady interactions of shock waves. Progress in Aerospace Sciences.

Read the latest blog to know the Top Five Reasons to Use Cradle CFD 2026.1