Ansys Fluent® is well-known in the commercial computational fluid dynamics (CFD) space and is praised for its versatility as a general-purpose solver. Its impressive breadth of physics capabilities makes finding a problem that Fluent cannot solve seemingly impossible. Fluent can be found in use across numerous industry segments globally, particularly in the automotive and aerospace sectors, as well as in energy, materials and chemical processing, and high-tech industries.
Ansys recently integrated support for AMD Instinct™ MI200 and MI300 accelerators into Fluent, its Flagship CFD Solver, significantly enhancing simulation efficiency and power. With the Fluent GPU Solver, simulations that once took weeks or months can now be completed in hours or days.
Director Jeremy McCaslin of the Ansys Fluids Product Management team adds, "Simulation is not a commodity anymore; it's a necessity. Fluent customers need to innovate designs faster, bring products to market more quickly, and do all of this more sustainably." He emphasizes the energy efficiency of GPUs, noting, "Large compute centers consume substantial energy, and using AMD Instinct GPUs can achieve similar computational tasks with significantly less energy."
As such, the release of Ansys Fluent 2024 R2 is a significant milestone. "This is our first version to have full AMD GPU support, enabling our users to run AMD Instinct GPUs out of the box, on MI200 and the MI300 series," says Ansys Distinguished Engineer Rongguang Jia.
McCaslin describes Ansys's long-term strategy with AMD GPU support, "We want to provide something that has all of your bases covered." By fully supporting AMD Instinct GPUs in the new release and strategically prioritizing feature development, Fluent software aims to deliver a powerful, comprehensive, and efficient solution. "Ensuring that GPU-supported features perform on par with or exceed customer expectations is crucial," McCaslin notes.
Travis Karr, AMD Corporate Vice President for HPC, adds, "We are pleased to closely partner with Ansys to integrate the industry's leading Fluent software into AMD's high-performance Instinct accelerators. Fluent users can now run AMD Instinct GPUs out of the box, increasing CFD efficiency and lowering energy consumption".
AMD and Ansys share a long history of collaboration
The collaboration between AMD and Ansys is rooted in a shared commitment to an open ecosystem. Ansys excels in simulation technology but relies on partners like AMD for hardware expertise. "AMD is a strategic partner within our overall philosophy of providing software solutions to our customers," says McCaslin. "We believe offering options like AMD Instinct GPUs is good for the overall pace of innovation in the market."
This partnership has long roots, with AMD contributing to Fluent's parallel performance and CPU core scaling capabilities for decades. Jia recalls, "When I started working at Ansys 20 years ago, AMD had just come out with the x86 64-bit, which transformed the industry." He continues, "Today, AMD Instinct GPUs power Frontier, the world's fastest supercomputer, offering substantial savings and contributing to corporate sustainability goals."
Supporting simulations that get larger and more complex by the day
Of course, the primary objective for many customers is the ability to perform simulations much faster and handle more complex tasks. "A high-fidelity simulation that once took weeks to complete can now be finished in just a day or two with heterogeneous (CPU+GPU) computing. Simulations that used to require several days are now done in the time it takes to have lunch, allowing for faster iterations and greater innovation," McCaslin shares. He notes that while some customers are excited to run faster simulations of existing models, others are excited to tackle larger and more detailed simulations. "Just a few years ago, running a billion cells felt really big. A billion cells seem smaller every day," he says.
To that end, integrating AMD Instinct GPUs with Fluent software offers significant advancements. McCaslin explains, "Fluent's coupled solver is very much the workhorse for customers doing steady state problems because it significantly improves the robustness and time to convergence. The high memory bandwidth and capacity of the AMD Instinct MI300X makes it a good choice for applications requiring steady-state analysis." He adds, "But, in general, all the models are well suited for AMD Instinct GPUs."
Karr agrees, “Ansys has effectively harnessed AMD Instinct GPUs' industry leading memory capacity and bandwidth to significantly accelerate complex simulations.
A standout feature of the AMD Instinct MI300A APU, is its shared GPU and CPU memory, which can allow logically complex models to benefit from continuing to run on AMD EPYC™ CPUs cores, while more computationally demanding tasks are handled by the GPU cores, without extensive data transfers. McCaslin says, "Having shared memory between CPU and GPU opens up a lot of interesting opportunities to get things onto the GPU where the compute will be faster." The ability to perform tasks on the CPU without data movement issues can accelerate the delivery of hybrid CPU-GPU solutions to customers more quickly.
McCaslin emphasizes the practical benefits of such developments, noting that GPUs could efficiently utilize large meshes created and iteratively modified by CPUs without significant data movement. "We don't want to be moving large amounts of data around just so that the GPUs can access the mesh that the CPUs created," he explains.
"We have customers with decades of established CPU CFD usage built around workflows developed over many years," says McCaslin. "If we can show a customer that's been running legacy workflows on CPU clusters for 15 years, 'Hey, look, I can get you that same outcome, but 10x faster, sometimes even more,' they'll make the jump," referring to adopting AMD Instinct GPUs and heterogenous workloads.
Porting for performance
Reflecting on porting components to HIP to support AMD Instinct GPUs in Ansys, Jia notes, "The AMD ROCm™ platform is progressing rapidly, and the HIP APIs used by our GPU Solver are comparable to CUDA APIs," adding, "despite differences, finding corresponding APIs are straightforward." Jia adds that the HIP compilers, built on Clang, have been "friendly to use."
Jia shared some impressive benchmarking results, saying, "For large or high-fidelity simulations, such as analyzing the impact of turbulence on the aerodynamic design of Formula 1 cars for example, we observed that one MI300X GPU is able to produce competitive performance for the coupled solver."
Figure 1 – HPC Competitive CFD Performance (AMD MI300X vs NVIDIA H100)1
Figure 2 – Formula 1 Race Car CFD Workload
Utilizing AI alongside Ansys Fluent with AMD Instinct GPUs
Ansys is leveraging AI capabilities alongside its core simulation tools. For example, Ansys SimAI is a new machine learning platform that enables engineers to leverage previous simulation data to assess the performance of new designs within a given design envelope—in minutes instead of hours. GPUs are crucial in enabling rapid execution of both high-fidelity simulations that define the design envelope, as well as AI-driven tasks that rapidly explore new designs within it. McCaslin emphasizes, "GPUs enable AI predictions to happen in seconds or minutes. But to explore entirely new design spaces, you need high-fidelity simulations to define the bounds of the design envelope—and running these simulations quickly on GPUs reduces the bottleneck for the overall design process." Rapid heterogeneous simulations facilitated by AMD Instinct GPUs make the nearly instantaneous AI results more impactful, ensuring an efficient overall workflow.
Moving forward starts with preserving the familiar
Ansys designed Fluent 2024 R2 to help users transition smoothly to GPU-enhanced workflows. Jia explains, "We make the transition easier by keeping the running environments exactly the same for our traditional CPU solver and the GPU solver."
McCaslin adds, "Selecting the GPU Solver is literally a checkbox in the Fluent launcher. Everything else is the same," he explains. For batch runs, users only need to add a "-gpu" flag to their command line. This accessibility ensures that users can leverage AMD Instinct GPU acceleration without learning new tools or significantly altering their workflows.
The benefits of GPU acceleration speak for themselves
The combination of enhanced performance and cost efficiency is a powerful driver for GPU adoption. As McCaslin says, "Performance, cost efficiency... it's a no-brainer." As more customers experience the benefits of GPU acceleration, he expects the shift towards GPU-enhanced workflows to continue accelerating.
Highlighting the overwhelmingly positive response, McCaslin says, "The user feedback around Fluent software running on CPUs AND GPUs is like, 'Wow, this is impressive. We can do the same workloads, but we can do them so much faster, mind-blowingly faster.'" Such an enthusiastic reception has only reinforced Ansys' commitment to prioritize GPU development for the Fluids product family.
Karr adds, “AMD is very excited by these results as well, providing the community access to the best technologies available. We look forward to continuing our strong partnership with Ansys, delivering best-in-class solutions to our users."
Find recommendations on installing Ansys Fluent on AMD Instinct™ systems at AMD’s Infinity Hub Github
Claims
1Testing performed by AMD May 2024 performance comparison between systems with 8 GPUs on Ansys Fluent 2024R2 Coupled Solver running the sedan_4m, aircraft_wing_14m, exhaust_system_33m, and f1_racecar_140m benchmark models. System configurations: 2xAMD EPYC™ 9554, 8x AMD Instinct™ MI300X-NPS1-SPX-192GB-750W, American Megatrends Inc. - 1.1 BIOS, NPS1 NUMA Config, 2TB memory, 877G storage, Ubuntu® 22.04.4 LTS, 113-MI3SRIOV-001 IFWI/vBIOS, ROCm™ 6.1.0-48. Versus 2x Intel Xeon 8480CL, 8x H100 80GB DGX 700W, Nvidia - 1.0.7 BIOS, NPS1 NUMA Config, 2TB memory, 1.8TB storage, Ubuntu® 22.04.4 LTS, 96.00.61.00.01 vBIOS, CUDA® 12.2. Results (Median):
MI300x - sedan_4m, 82.67
H100 - sedan_4, 76.79
MI300x - aircraft_wing_14m, 71.01
H100 - aircraft_wing_14m, 67.67
MI300x -exHaust_system_33m, 128.24
h100 - exhaust_system_33m, 138.46
MI300x - f1_racecar_140m, 115.76
H100 f1_racecar_140m, 105.63
Server manufacturers may vary configurations, yielding different results. Performance may vary based on use of latest drivers and optimizations. MI300-58