The Pulse
Convergent Science Completes Roughly 536-Million-Cell CFD Simulation in 3.62 Hours on B200 GPUs
Convergent Science completed a roughly 536-million-cell computational fluid dynamics simulation in 3.62 hours on two NVIDIA HGX B200 instances. The joint study also measured how CONVERGE CFD software scaled across RTX PRO 6000 Blackwell Ser
AI.info Team · Written with AI · fact-checked ·
Convergent Science completed a roughly 536-million-cell computational fluid dynamics (CFD) simulation in 3.62 hours on two NVIDIA HGX B200 instances, using 16 NVIDIA Blackwell GPUs. The run tested how far the configuration could carry a case; it was separate from the study’s scaling matrix. Convergent Science and CoreWeave used CONVERGE CFD software for the project, which also measured GPU solver accuracy against a well-established test case and examined scaling across two hardware configurations.
Why the test used direct numerical simulation
The team chose direct numerical simulation (DNS), which computes turbulent flow across its scales without a turbulence model. CoreWeave describes DNS as the highest-fidelity simulation approach and a reference for calibrating production turbulence models. Resolving the smallest eddies requires very small cells and short time steps, making DNS computationally demanding.
Many industrial simulations instead use Reynolds-Averaged Navier-Stokes (RANS) methods. These use a turbulence model to compute large motions and estimate smaller ones, reducing the computational load. CoreWeave notes that DNS carries fewer equations than the RANS models used in everyday industry work. A production case with the same cell count may therefore need more memory, particularly if it includes additional physics such as combustion reactions.
Two GPU systems, two kinds of scaling
The scaling matrix used a validated three-dimensional case of turbulent flow losing energy over time. In weak scaling, the case grows as GPU resources increase; in strong scaling, the case stays fixed while more GPUs work on it. CoreWeave leads with weak scaling because engineering teams often need to run larger, more detailed cases. Strong scaling tests how much sooner an existing case can finish.
On NVIDIA RTX PRO 6000 Blackwell Server systems, the weak-scaling runs grew from about 16.8 million cells on two GPUs to about 67.1 million on eight. The strong-scaling tests held the case at about 67.1 million cells and used two, four or eight GPUs. On NVIDIA HGX B200, weak-scaling runs ranged from about 33.6 million cells on one GPU to about 134.2 million on four. Strong scaling tested about 268.4 million cells across eight, 12 and 16 GPUs.
The two systems differ in memory capacity and GPU connections. The RTX PRO configuration has 96 GB of memory per GPU and PCIe Gen5 connections. The HGX B200 configuration has 180 GB per GPU and fifth-generation NVLink within a node. For communication between nodes, the test used a 400 Gb/s NVIDIA Quantum-2 InfiniBand network. CoreWeave says that communication between GPUs can affect scaling because GPUs exchange boundary data during each time step.
Memory sets the limit before speed
Cell count alone does not determine how much memory a simulation needs. Requirements also depend on the physics being modeled, the precision of stored values, the solver’s working data and whether the mesh changes during the run. The tested case used single-phase DNS without reactions. CoreWeave says adaptive mesh refinement can concentrate cells where needed and remove them elsewhere, so the cell count varies during a run and capacity should account for the peak.
The 536-million-cell run establishes the largest case this configuration and hardware carried. The separate scaling runs measure performance as GPU counts or case sizes change. CoreWeave says the full study includes validation against published DNS data, weak and strong scaling on both GPU configurations, and analysis of efficiency limits in a double-precision solver. The jobs ran on bare-metal nodes through SUNK, CoreWeave’s implementation of Slurm on Kubernetes.
“This was one of the best clusters we've tested CONVERGE on.”
Cooper Burns, HPC Team Lead at Convergent Science