Ansys HPC Packs for GPUs

Ansys HPC Packs don’t just define licensing costs for CPUs; they also define the cost of running Ansys on GPUs. Many Ansys simulation applications are still CPU-reliant, but GPUs are paramount in modern high-performance computing. Some Ansys simulation solvers, including Fluent, HFSS SBR+, and Mechanical ADPL, are accelerated or native to GPUs (with certain solvers requiring FP64-capable GPUs).

GPU microarchitecture differs from CPUs, with hundreds of times more cores than those on CPUs. So instead of using core counts as the variable, they use Streaming Multiprocessors, or SMs to count HPC Packs. Not to go into full detail about GPU microarchitecture, but SMs are the building blocks of a GPU that store multiple cores, cache, and controllers. Think of SMs as a group of workers tasked with computations, memory management, instruction pipelines, and a GPU is a factory of these groups. These SMs enable parallelized computing, thus are what Ansys counts as “cores” in this case.

HPC Workgroups are defined just by the cores enabled and can be split however you like, as long as you already have the application license. These are defined by the SMs or cores enabled so it's not as confusing as HPC Packs. For each increment of workgroup, you can split your workload however: an organization with 32 HPC workgroup can be split 8-8-8-8, or 24-8, or 16-16. HPC Workgroups are more flexible in multi-system deployments.

SMs	HPC workgroup	HPC Packs
1-40	0	0
41-48	1-8	1
49-72	9-32	2
73-168	33-128	3
169-552	129-512	4
553-2088	513-2048	5

Finding a GPU's SM count is not as easy as looking at the spec sheet. We listed current-gen and popular GPU models and their SM counts found by scouring GPU microarchitecture whitepapers and third-party sources like techpowerup.com. We will also compare how many HPC packs would equate to each GPU. We also added the GPU memory to help gauge which GPUs suit your potential model size. (This does not affect the HPC pack licensing.)

GPU	SM Count	HPC Packs Needed	GPU Memory
NVIDIA RTX 4500 Ada	60	2	24GB GDDR6
NVIDIA RTX 5000 Ada	100	3	32GB GDDR6
NVIDIA RTX 6000 Ada	142	3	48GB GDDR6
NVIDIA A800 40GB Active	108	3	40GB HBM2e
NVIDIA H200 NVL	132	3	141GB HBM3e
NVIDIA RTX PRO 6000 Blackwell	188	4	96GB GDDR7

Since SMs cannot be disabled at will like with CPU cores, if the GPU configuration exceeds the SM count threshold for HPC packs, additional purchases of HPC packs are required. Therefore, depending on the configuration, careful consideration of single or multiple GPUs is necessary for optimizing each HPC pack license. Some GPUs should be prioritized over others on this list since they maximize the number of available SMs per HPC Pack.

For example, a single NVIDIA RTX 5000 Ada (100 SMs) requires 3 HPC packs. The faster and larger NVIDIA RTX 6000 Ada (142 SMs) would also cost 3 HPC packs. By paying the price of the higher-tier card, you can maximize the price you’re paying for Ansys HPC pack licensing while getting better performance. But for multi-GPU configurations, the story is a little different. In the table below, we detail the number of HPC packs needed to run multiple GPUs:

GPU	SM Count	HPC Packs Needed	GPU Memory
1x NVIDIA RTX 5000 Ada	100	3	32GB GDDR6
2x NVIDIA RTX 5000 Ada	200	4	64GB GDDR6
4x NVIDIA RTX 5000 Ada	400	4	128GB GDDR6
1x NVIDIA RTX 6000 Ada	142	3	48GB GDDR6
2x NVIDIA RTX 6000 Ada	284	4	96GB GDDR6
4x NVIDIA RTX 6000 Ada	568	5	192GB GDDR6
1x NVIDIA A800	108	3	40GB HBM2e
2x NVIDIA A800	216	4	80GB HBM2e
4x NVIDIA A800	432	4	160GB HBM2e
1x NVIDIA H200 NVL	132	3	141GB HBM3e
2x NVIDIA H200 NVL	264	4	282GB HBM3e
4x NVIDIA H200 NVL	528	5	564GB HBM3e
1x NVIDIA RTX PRO 6000 Blackwell	188	4	96GB GDDR7
2x NVIDIA RTX PRO 6000 Blackwell	376	4	192GB GDDR7
4x NVIDIA RTX PRO 6000 Blackwell	752	5	384GB GDDR7

In a 4x GPU configuration scenario, consider this:

4x NVIDIA RTX 6000 Ada needs 5 HPC Packs to run it, versus the 4 HPC packs needed for 4x NVIDIA RTX 5000 Ada, which may sway some purchasing decisions. Be strategic with licensing by maximizing per core or per SM limit without spilling over to the next bracket. The NVIDIA RTX 5000 Ada may be a bit slower, but if the simulation size calls for 4x GPUs, purchasing an extra GPU without the need for an extra license can influence budget and performance considerations.
4x NVIDIA RTX PRO 6000 Blackwell would be a better purchase if the budget allows for 5 HPC Packs, since it has improved throughput and 96GB of memory per GPU. It can also be fitted into a workstation as opposed to the H200 NVL (a server-only card). The only downside to this approach is the lack of native FP64 double precision throughput.
For 4x GPU configurations that require double precision FP64, there are two options: workstation or server. The NVIDIA H200 NVL is a server-only GPU that features the highest memory capacity, bandwidth, and throughput, making this option the best GPU to use for compute-only applications (as this GPU does not have video out). The NVIDIA A800 40GB Active requires 4x HPC packs and has native double-precision FP64 that can be outfitted in a workstation with display out, but delivers less memory and throughput than the alternative. We would recommend going with a GPU server with H200 NVLs if FP64 is necessary.

This whole HPC Pack and GPU SM count consideration is challenging to navigate. If you have any questions, our Exxact engineers are more than happy to assist with any decision-making you may need.

All Software

No search results match your search query.