Ansys Fluent GPU Benchmark with DRD & Exxact

More hardware doesn’t always equate to better performance. In collaboration with DRD, leading Ansys software vendor and consultant, our team at Exxact provided the hardware to run GPU benchmarks on Ansys Fluent 2026. Exxact is a hardware partner for DRD and DRD customers, offering configurable workstations and servers for Ansys worflows.

We ran a series of benchmarks on Ansys Fluent using a 41.3 million element exhaust model to examine how different GPU configurations, GPU generations, and precision settings affect both solve time and memory usage.

We tested one, two, four, and eight GPU configurations on

• NVIDIA RTX PRO 5000 Blackwell (48GB GDDR7 VRAM)
• NVIDIA RTX PRO 6000 Blackwell (96GB GDDR7 VRAM)
• NVIDIA H200 NVL (141GB HBM3e VRAM) with and without NVLink
• All cases were run on an identical 41.3M element exhaust model using double precision measured in time to completion (seconds)
• DRD ran a 1:1 CPU Core-to-GPU ratio was maintained throughout
• Runs that exceeded available VRAM (i.e. did not fit in GPU memory) were excluded

Here's what we found, and what it means for your hardware decisions.

Ansys Fluent CPU vs GPU Takeaways

You might’ve heard it before, GPUs are a necessity for Ansys Fluent and CFD workflows. For teams evaluating whether to invest in GPU hardware at all, here’s a very convincing story.

• A single H200 NVL (658.68s) is more than 15x faster than 32 CPU cores
• 8x RTX PRO 5000 (299.80s) is roughly 33x faster than the 32 CPU cores.

*NVIDIA H200 NVL performance plateaus due to communication overhead and the model not saturating all active GPUs

It's also worth noting that optimal CPU-only performance for Fluent typically occurs at around 0.5–2 million cells per core. At 41.3M elements, even 32 cores sit at the upper end of that range, meaning the CPU numbers here reflect a reasonably well-tuned configuration if this is your own option for now. But for CFD, we strongly recommend investing in GPU hardware

Ansys Fluent GPU Benchmark Takeaways

The headline numbers across the three GPU configurations tested — NVIDIA RTX PRO 5000, RTX PRO 6000, and H200 NVL — show GPU scaling has a ceiling, depending on the test case.

• Going from 4 to 6 to 8 GPUs produces increasingly smaller gains, compared to going from 1 to 2 to 4.
  • From 4 GPUs to 6 GPUs cuts solve time by roughly 71 seconds (15.8% speedup).
  • From 6 GPUs to 8 GPUs cuts solve time by roughly 80 seconds (21.1% speedup).
• There's a meaningful gap between the NVIDIA RTX PRO 5000 and RTX PRO 6000, but there are diminishing returns on scaling multiple GPUs at this element count.
  • At 4 GPUs, the RTX PRO 6000 is about 15% faster — a very sizable performance gain.
  • By 6 GPUs, that gap narrows to just 32 seconds, roughly an 8.5% difference.
  • With 8 GPUs the difference is under 5 seconds or less than 2%.
• NVLink at this 41.3M element count shows mixed and unexpected characteristics. The NVIDIA H200 NVL is overkill for this size of simulation.
  • At 2x and 4x GPU configurations, the NVIDIA H200 NVL with no NVLink (390.90s) outperformed the NVIDIA H200 NVL with NVLink (436.13s).
  • At 6x GPU both GPUs hit their plateau with H200 NVL with NVLink (331.54s) outperforms the H200 NVL no NVLink (419.58). Any additional GPUs past 6x for H200 NVL won't contribute any more performance.
  • All the NVIDIA H200 NVL numbers aren’t remarkable compared to NVIDIA RTX PRO 5000 and NVIDIA RTX PRO 6000 numbers at this element count. In 2x GPU, H200 NVL is only 8% faster than RTX PRO 6000 and 2% slower in 4x GPU.

Once you account for the cost and power draw of two additional GPUs, the key insight here is not that GPU scaling doesn't work, it certainly does. The simulation size has to be large enough to justify the parallelism. At 41.3M elements, this benchmark is likely brushing up against the lower end of where 6x and 8x GPU configurations can truly shine. Larger meshes in the 60–100M+ element range would show more decisive scaling benefits at higher GPU counts.

Precision Settings and Ansys’s Hybrid Precision for VRAM-Constrained Configurations

One underappreciated tool in Fluent's configuration arsenal is precision mode. If your model is pushing the VRAM limits of a 4x GPU configuration, switching from double to hybrid or single precision may let you run the same problem without adding GPUs.

Hybrid Precision is a newer Ansys Fluent feature that uses single precision for field variables and double precision for AMG inner cycles. It retains better numerical accuracy than pure single precision.

The benchmark tested three settings:

• Single precision cuts VRAM usage by 33% compared to double precision and improves solve time by 15.6%
• Hybrid precision cuts VRAM usage by 9% compared to double precision, and improves solve time by 11.5%.

Ansys Fluent Hardware Configuration Recommendations

Based on these findings, here's how to think about GPU selection for Fluent workloads in this size range:

At Exxact, we deliver high-performance workstations and servers for engineers running their most demanding simulation workloads. In partnership with DRD — our industry expert for all things Ansys — we help teams find the right hardware for the job.

Our engineers are ready to help you configure a GPU solution that fits your need; talk to our team about running a test drive on your own dataset. Any questions regarding performance on a GPU with your simulation workflow? DRD has amazing engineers that can point you in the right direction!