GPU for DEM and SPH simulation: sizing memory, FP64 and SM licences for Ansys Rocky

Which GPU for Ansys Rocky

For Ansys Rocky the particle shape decides the card. Simulations with shaped particles, polyhedra for example, need strong FP64, which among the cards we supply means the NVIDIA H200 NVL; sphere-only cases depend on memory bandwidth and SPH on single-precision speed and bandwidth, where the NVIDIA RTX PRO 6000 Blackwell in its Workstation or Server Edition fits. After that, GPU memory sets how many particles fit, and Rocky's licence counts the card's streaming multiprocessors.

DEM, the discrete element method, follows every particle of a bulk material and its contacts, which is how mixers, chutes, mills and conveyors are simulated. SPH, smoothed particle hydrodynamics, represents a fluid itself as particles, used for splashing, wading or free-surface flows. Rocky does both on GPUs, and the figures here come from Ansys's "Rocky GPU buying guide" on Ansys Innovation Space, marked "Last Updated: April/2026", which we read on 25 October 2026.

The guide is explicit about priorities. For shaped particles it advises prioritising "higher double-precision performance"; for spheres only, higher memory bandwidth and, for very large cases, more memory; for SPH, high single-precision performance, memory bandwidth and enough memory for millions of elements. It adds that cards with weak double precision but good bandwidth do well on spheres and poorly on shaped particles. Where SPH and DEM run together, the bottleneck can be either one, depending on particle counts and shapes.

Rocky DEM GPU requirements: memory per million particles

Rocky's 2026 R1 benchmarks report total GPU memory for typical cases, run on H100, A100 and L40 cards. We divided each figure by the particle count to get a planning value per million, and from that the particles that fit on 96 GB and 141 GB.

Ansys Innovation Space, "Rocky GPU buying guide", last updated April 2026, 2026 R1 benchmarks (rotating drum, cyclone separator with one-way Fluent coupling, car wading), read 25 October 2026. Per-million values and particle counts are our arithmetic from nominal card memory, rounded down, without headroom.

Two things follow. The memory per particle depends on the case, not only the particle type: the spheres of the cyclone take less than a third of the memory per million of the spheres in the drum. Rocky itself writes that the 90 GB of its 32 million polyhedra could hold about 185 million spheres, close to the cyclone and far from the drum. Size from a case of your own, run on any GPU, and scale the measured memory to the target count. The second point is coupling: in the cyclone, multi-GPU runs used about 92 GB against about 77 GB on one GPU, 17 per cent more, so a coupled case split over cards needs that margin. For SPH, the car wading case with 16 million elements peaked at about 27 GB across two GPUs.

The minimum requirements in the guide are at least 4 GB of GPU memory, fast double precision, CUDA compute capability 6.0 or higher and a driver that supports CUDA 12.8 or higher. RTX PRO Blackwell cards have compute capability 12.0. The guide's recommended lists name the A30, A100, L40, H100, H100 NVL and H200 among server cards and Ampere-generation RTX cards among workstation cards, and calls the A100 "the Rocky team's preferred choice". No Blackwell RTX PRO card is named. They meet the memory, compute capability and driver minimums; whether their FP64 rate counts as "fast" in the guide's sense is for Ansys to confirm, which is why we place them on sphere and SPH cases, not on shaped particles.

FP64, FP32 and bandwidth of the cards we supply

NVIDIA H200 product page and RTX PRO product pages and datasheets; RTX PRO 6000 FP64 is our arithmetic from the 1/64 rate in NVIDIA's RTX Blackwell PRO architecture whitepaper, since the product pages list none.

Read against Rocky's priorities, the table gives a clear split. For shaped particles the H200 NVL has about 16 times the FP64 rate of an RTX PRO 6000, by our arithmetic, and three times the bandwidth. For sphere-only cases the H200 NVL's 4.8 TB/s is the higher bandwidth, while the RTX PRO 6000 holds the same order of particles per card and, in its Workstation Edition, also drives the screens for setup and post-processing. For SPH the RTX PRO 6000 has twice the FP32 rate and the H200 NVL three times the bandwidth; we found no published Rocky result that sets the two cards against each other, so a test run of one of your SPH cases on both is the way to decide. Our FP64 GPU server comparison explains the 1/64 rate, and our comparison of GPU memory bandwidth covers the other cards.

We supply both card families, as cards or in workstations and servers built for Rocky; send us your particle counts, shapes and whether you couple with Fluent, and our team replies with a configuration.

Rocky multi-GPU licences, counted by streaming multiprocessors

Rocky does not license per card. The guide states that licensing "is based on the total number of SM's across all GPU's irrespective of the number of GPU's", that all SMs of a card are used, that the base licence covers one job with up to 112 SMs, and that one Rocky HPC task enables 14 more; an HPC-8 licence holds 8 tasks, 112 SMs. NVIDIA's whitepaper lists 188 SMs for the RTX PRO 6000 Blackwell Workstation and Max-Q Editions. The Server Edition has the same 24,064 CUDA cores, which at 128 cores per SM gives 188 by our division.

Licence rules from the Rocky GPU buying guide (April 2026); SM counts from NVIDIA's RTX Blackwell PRO whitepaper (RTX PRO 6000) and, for the RTX PRO 5000, our division of NVIDIA's 14,080 CUDA cores by 128 per SM; tasks and licences are our arithmetic, rounded up. Confirm the count with your Ansys channel.

The RTX PRO 5000, in its 48 or 72 GB version, is by this arithmetic just inside the base licence, which makes it a sensible single-card Rocky workstation when the cases fit its memory. The SM count of the H200 NVL is not listed on NVIDIA's H200 product page or in the H200 NVL product brief we opened, so ask your Ansys channel for its licence count before you compare a server of H200 NVL cards with RTX PRO 6000 cards. The licence count, the card's FP64 rate and its memory together decide which card suits a case, which is why we size from your cases.

One server or several

Rocky runs several GPUs in one machine within the licence rules above. Across machines the guide is cautious: since 2026 R1 distributed GPU computing is a beta that applies only to spheres without coupling, on Linux, and the guide advises keeping all GPU cards in a single server because distributed runs are not fully supported. A team that expects to grow should therefore plan the chassis for the final card count from the start. Our article on how many GPUs fit in one server explains the slot, power and airflow limits.

On speed, Rocky's benchmarks compare GPUs with a 48-core Intel Xeon Gold 6542Y: up to 41 times faster for DEM with several GPUs, about 18 times for CFD-DEM on one GPU and 31 times on several, and for SPH about 23 times on one GPU in the worst case and about 65 times on two. Those runs used H100, A100 and L40 cards, not the cards in our tables.

Our reading of the Rocky GPU buying guide. Wall power is our estimate (cards at 600 W, two AMD EPYC 9355, plus 15 per cent), as in our four-card and H200 NVL server guides, and is confirmed by a measurement on the delivered system.

For CFD-DEM, Rocky couples with Ansys Fluent, whose GPU solver follows its own memory and SM licence rules; our CFD simulation guide sets them out. For shaped-particle work, our guides to the two-card and four-card H200 NVL server cover chassis and power, and for SPH the four-card RTX PRO 6000 Server Edition server. Process engineering departments that also plan AI work will find the wider picture in our article on AI servers for manufacturing companies. If you are unsure which card your mix of cases needs, describe the cases and the licence you hold and we size the card count to it.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 6000 in its Workstation, Max-Q and Server Editions, the RTX PRO 5000 and the H200 NVL as cards, and build workstations and servers to order, assembled and tested before delivery, on one EU contract and invoice, with manufacturer warranty on every component and the assembly covered by us. Operating system and driver are installed on request. Quantities and lead time are confirmed in writing with the quote.

FAQ