Moving Towards 1000x Productivity: Introducing Hexagon's scFAST, a fully NVIDIA GPU-Optimized CFD Solver

Since its inception in 1984, Software Cradle has been at the forefront of computational fluid dynamics (CFD) innovation. Now part of Hexagon (following its integration into MSC Software in 2019), Cradle offers industry-leading solvers: scFLOW, designed for general fluid flow simulations using an unstructured polyhedral mesh, and scSTREAM, a fast Cartesian-mesh solver optimized for electronics cooling and thermal management.

In our continuous drive for performance and efficiency, we have introduced scFAST, a fully GPU-optimized CFD solver. Built specifically for NVIDIA Hopper and NVIDIA Blackwell architectures, scFAST marks a significant leap in simulation speed and scalability.

Why GPU Optimization Matters

Traditional CFD solvers rely on CPUs, but GPUs offer massive parallel processing capabilities that can dramatically accelerate simulations. scFAST is built on the foundation of scFLOW but is re-engineered to run entirely on GPUs, eliminating CPU bottlenecks and optimizing memory efficiency.

Key Innovations in scFAST

1. GPU-Resident Architecture for Maximum Speed

scFAST operates entirely on the GPU, keeping all data in GPU memory. This eliminates the frequent data transfers between the CPU and GPU, significantly reducing latency and improving overall simulation speed.

2. Reduced Memory Usage with the RM-MCGS Method

One of scFAST's breakthroughs is the Reduced-Memory Multicolor Gauss-Seidel (RM-MCGS) method. Instead of storing large Jacobian matrices, RM-MCGS leverages the Fréchet derivative, drastically reducing memory consumption. This is especially beneficial for GPUs, which thrive on computational intensity but have limited memory compared to traditional CPU clusters.

3. Hardware Adaptability with Portable OpenMP Code

scFAST is designed with portable OpenMP code, allowing it to run efficiently across both multicore CPUs and GPUs. This makes it flexible and adaptable to different hardware configurations, ensuring seamless integration into existing workflows.

4. Code Optimization Using C++ Templates

Using C++ templates, scFAST generates highly optimized code paths based on specific user settings. This ensures that performance is maximized for different simulation scenarios, delivering unparalleled efficiency.

Performance and Efficiency: The 450x Leap
The results speak for themselves. On a single NVIDIA H100 GPU, scFAST delivers 15% faster performance than the previous-generation solver running on 384 CPU threads (four AMD EPYC 7763 processors). This translates to an overall 450x performance improvement, enabling engineers to run high-fidelity simulations in a fraction of the time previously required.

     

Figure 1: Speedup using 53M elements mesh for the Onera OM6 wing. Optional RM-MCGS (Reduced Memory) algorithm saves memory when needed at the cost of some speed.

Handling Large-Scale Problems on a Single GPU

One of scFAST's most revolutionary capabilities is its ability to handle massive simulations on a single GPU. With support for over 50 million elements per GPU, scFAST enables single-GPU simulations that previously required multi-node CPU clusters. This breakthrough is particularly transformative for industries like aerospace, automotive, and energy, where large-scale CFD problems are the norm.

 

Figure 2: CRM-HL simulation results using a 68-million-element mesh on a single NVIDIA A100 GPU.

The Future: 1000x Performance Gains & AI Integration

"We are excited to provide an early release of scFAST to our CFD customers," says Subham Sett, VP of R&D and Innovation, Design & Engineering, Hexagon. "A wider rollout is planned later this year, with full compatibility for NVIDIA Blackwell architecture."

Beyond this, Hexagon is investing in faster meshing, AI-driven reduced-order models, and enhanced CAD integration. With these advancements, we are confident in achieving 1000x productivity gains, making high-performance CFD simulations more accessible, efficient, and accurate than ever before.

Stay tuned for updates as we continue to push the boundaries of CFD performance!