GPU server for chemistry and materials research: molecular dynamics, DFT, FP64, lab LLMs

Which GPU server a chemistry lab needs

A GPU server for chemistry and materials research is chosen by the precision of the codes it runs. Molecular dynamics in GROMACS, AMBER and LAMMPS runs in single or mixed precision and fits servers with the NVIDIA RTX PRO 6000 Blackwell Server Edition, while plane-wave DFT in VASP and other double-precision codes belong on the NVIDIA H200 NVL with 30 TFLOPS of FP64. Many labs run both kinds of work, plus training of machine-learned potentials and a language model over their own documents, and end up with two card types in one or two servers.

For a director the picture is simple. A research lab in coatings, batteries, catalysis, polymers or flavours and fragrances runs simulations for days at a time, and each card runs one or a few of them. More cards mean more simulations in parallel, more candidate molecules screened per week and shorter waits for the team. The card type decides whether the double-precision part of the work runs at full speed. Our guides cover the neighbouring fields: CFD and engineering simulation, pharma and life sciences with AlphaFold 3 and BioNeMo, and a shared university lab server with Slurm and quotas.

Single, mixed or double precision: what the codes document

The table summarises what each project states about precision and multi-GPU use, as read on 24 October 2026.

GROMACS 2026.4 installation and performance guides; AMBER GPU hardware and GPU support pages (last modified 7 September 2023 and 6 June 2024); LAMMPS GPU package page (release of 30 September 2026); VASP wiki page on the OpenACC GPU port; MACE README on GitHub; all read on 24 October 2026. Card fit is our reading of these documents.

GROMACS. The installation guide for GROMACS 2026.4 states that simulations "are normally run in 'mixed' floating-point precision" and describes a double-precision build as "slower, and not normally useful". The performance guide repeats the point: "Do not use double precision unless you are absolutely sure you need it." Non-bonded forces, and optionally PME, bonded forces, update and constraints, can be offloaded to the GPU, and the guide recommends one to three ranks per GPU, typically one rank per GPU in GPU-resident mode. The build needs CUDA 12.1 or newer and compute capability 5.0 or higher, which every card in this article meets.

AMBER. AMBER's GPU page describes pmemd.cuda's precision model as a "custom designed hybrid single / double / fixed precision model termed SPFP". On scaling it is blunt: "Scaling over multiple GPUs within a single node is not really feasible anymore for PME calculations", with an exception for implicit-solvent runs above 2,500 atoms on cards that talk peer to peer. A second AMBER page says the benefit of running one simulation on two or more GPUs is "marginal", except for replica-exchange runs, and recommends independent simulations on separate GPUs. That is how a four-card server is used for AMBER: four simulations, one per card.

LAMMPS. The GPU package page (release of 30 September 2026) says it "allows for GPU computations to be performed in single or double precision, or in mixed-mode precision", where mixed mode computes pairwise forces in single precision and accumulates them in double. It advises users to test whether single or mixed precision gives accurate results for their system. Only part of PPPM runs on the GPU, and the FFT stays on the CPU; bonded terms also run on the CPU. The page says that 2 to 10 MPI tasks per GPU "will often give the best performance", so a LAMMPS server needs more processor cores than NVIDIA's minimum of six per GPU.

Tell us which codes and versions your team runs, with a typical system size, and our team replies with a configuration that matches their precision.

DFT and quantum chemistry: why FP64 decides the card

NVIDIA's RTX Blackwell PRO architecture whitepaper (v1.0) states: "The FP64 TFLOP rate is 1/64th the TFLOP rate of FP32 operations." The chip carries two FP64 cores per SM, which "are included to ensure any programs with FP64 code operate correctly". On the RTX PRO 6000 Server Edition, rated at 120 TFLOPS of FP32, that is 1.9 TFLOPS of FP64 by our arithmetic. The H200 NVL is a Hopper card with native FP64 units, which NVIDIA rates at 30 TFLOPS of FP64 and 60 TFLOPS on its Tensor Cores.

NVIDIA H200 product page and RTX PRO 6000 Blackwell Server Edition product page; FP64 rate from NVIDIA's RTX Blackwell PRO architecture whitepaper v1.0, read on 24 October 2026. The RTX PRO 6000 FP64 figure is our arithmetic (120 ÷ 64), since NVIDIA's product page lists none.

VASP is the clearest case. Its wiki page on the OpenACC GPU port, released with VASP 6.2.0, lists the NVIDIA data-centre GPUs it has tested, from the P100 to the H100, the Hopper generation of the H200 NVL. Running VASP on other NVIDIA GPUs, "gaming" hardware for example, "is technically possible but not advisable", because "these GPUs are not well suited since they do not offer fast double-precision floating-point arithmetic (FP64) performance". The RTX PRO 6000 Server Edition is not on the tested list, and by our reading its FP64 rate of 1/64 of FP32 is exactly the property that sentence describes. The same page states that the GPU version runs "a single MPI-rank per available GPU" and that NCCL "has large performance benefits in the majority of cases". On H200 NVL cards joined by an NVLink bridge, traffic between the GPUs can use the NVLink bridge at 900 GB/s per GPU instead of PCIe.

For other electronic-structure codes the rule is the same: ask the developers whether the GPU kernels compute in FP64. A 16-fold difference in FP64 throughput outweighs the RTX PRO 6000 Server Edition's lead in FP32. Our article on FP64 GPU servers compares the cards that still carry full double precision.

Machine-learned potentials and molecular ML models

Machine-learned interatomic potentials sit between the two kinds of work: they are trained on DFT data and then drive molecular dynamics at a fraction of the cost. The MACE README states that "the default is float64 but float32 gives a significant speed-up (usually a factor of x2 in training)", and that multi-GPU training uses PyTorch's DistributedDataParallel with the --distributed flag. The README does not compare the accuracy of the two settings.

That leaves a decision for the lab. A team that has validated float32 for its systems trains on RTX PRO 6000 Server Edition cards, where FP32 is the strong suit. A team that stays with float64 trains on the H200 NVL, because 1.9 TFLOPS of FP64 would slow every epoch on an RTX PRO card. Other molecular ML models, for property prediction or generation, follow the same test: the precision in which they are trained and validated. Our fine-tuning server guide covers the memory and multi-GPU side of training, and the H200 NVL vs RTX PRO 6000 comparison for fine-tuning sets the tensor rates side by side.

A language model on lab documents

Most labs also want an assistant that searches their papers, protocols, safety data sheets and project reports. By our users-per-card rule, 0.9 × driver-visible memory − 3 GiB per card − weights, one RTX PRO 6000 Server Edition offers 83.0 GiB and one H200 NVL 123.4 GiB.

Our estimate by the users-per-card rule: gpt-oss-120b weights 60.8 GiB and 1.125 GiB of 16-bit KV cache per 32K conversation, Qwen3.8-27B FP8 28.8 GiB and 2.14 GiB, as in our two-card server guide.

A lab of fifty people rarely has more than a handful of requests in flight, so one card or a MIG instance often carries the assistant, and the other cards keep running simulations. The sizing of larger assistants is in our guide to a two-card RTX PRO 6000 Server Edition server. The retrieval platform, document access rights and the query log are part of our Private AI/ML service.

Four configurations for a chemistry or materials lab

These start from the reference configurations on our AI servers page. Memory and processor cores follow NVIDIA's configuration guide for NVIDIA-Certified Systems: at least twice the GPU memory in system memory, at least six cores per GPU and two sockets.

Reference configurations of our AI servers page and the NVIDIA-Certified Systems configuration guide (30 September 2026). Wall power is our estimate: cards at 600 W plus two processors (EPYC 9135 at 200 W for two RTX PRO 6000 Server Edition, EPYC 9355 at 280 W for two H200 NVL and for four cards of either kind), plus 15 per cent for the rest.

A lab that runs both MD and DFT gets two servers: four RTX PRO 6000 Server Edition cards for simulation throughput and two or four H200 NVL for FP64. For large DFT campaigns or training of foundation potentials, our training node carries eight H200 NVL with NVLink bridges, 2.3 TB of DDR5 ECC, eight NVMe drives and 400G networking; our page on the four-card H200 NVL server shows what one NVLink domain holds. The card choice in general is in our comparison of four RTX PRO 6000 or four H200 NVL.

Processors, memory and storage around the cards

Simulation codes use the processors more than LLM serving does. LAMMPS runs 2 to 10 MPI tasks per GPU and keeps the PPPM FFT and the bonded terms on the CPU, and GROMACS splits the CPU cores among its ranks. Four cards with LAMMPS at up to ten tasks each call for up to 40 cores, well above the 24 that NVIDIA's guide asks for four GPUs. We therefore size the processors from the codes; our guide to EPYC or Xeon for a GPU server explains the trade-offs.

Trajectories set the storage. An uncompressed frame of one million atoms holds 12 MB of single-precision coordinates, so 10,000 frames come to 120 GB by our arithmetic; compressed formats such as GROMACS's XTC store much less. Local NVMe drives take the running jobs, while finished trajectories move to the lab's file server or object storage. Our guide to NVMe storage for a GPU server covers drive classes and endurance.

Power, cooling and continuous load

A simulation keeps its card near full power for days, unlike a chat service that idles between requests. By our estimate a two-card server with two EPYC 9135 draws about 1.8 kW at the wall and a four-card server with two EPYC 9355 about 3.4 kW, with cards at 600 W. With H200 NVL at 600 W, a two-card server with two EPYC 9355, as in our H200 NVL configurations, draws about 2.0 kW and a four-card server about 3.4 kW. We plan a 230 V 16 A feed at 80 per cent, 2.94 kW, so the four-card server needs a 32 A single-phase or a three-phase feed. Dell's technical guide for the PowerEdge XE7740 limits the inlet temperature to 30 °C with H200 NVL cards, so the room counts as much as the feed. Our guide to GPU server power on 16 A, 32 A and three-phase works through redundant supplies.

How we build and deliver the lab server

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. Sizing is free. We assemble the server, test it before delivery and deliver it anywhere in the EU on one contract and invoice, with manufacturer warranty on every component and the assembly covered by us; other destinations go through our logistics partner. On request we install the operating system, NVIDIA driver, CUDA and a container runtime, and commission the server on site. Quantities and lead time are confirmed in writing with the quote.

For a quote we need the codes and versions, the precision each runs in, typical system sizes, how many simulations run at once, any training of potentials and the rack's power feed; send these through the form, and we reply with a configuration and a written quote.

FAQ