No products in the cart.
How to Size a Molecular Dynamics Workstation: GPU, CPU, and MIG
Most molecular dynamics hardware guides start with the GPU and stop there. That gets the order wrong. For a lab running AMBER, GROMACS, NAMD, or LAMMPS alongside structure prediction, three questions determine the build, and only the first is about GPU memory:
- Does the job fit? A VRAM ceiling is absolute — no scheduling change fixes an out-of-memory error.
- How are you parallelizing? One large trajectory and fifty replicas want opposite hardware.
- Where is the actual bottleneck? In several engines the answer is the CPU, and a faster GPU changes nothing.
This guide works through all three. It also corrects a specification error circulating in vendor comparisons right now that pushes labs toward more hardware than they need.
Question 1: Does the job fit?
Here is the thing most MD hardware guides get backwards. For classical molecular dynamics, VRAM is usually not the binding constraint on a modern card. Even large explicit-solvent systems are comparatively modest in GPU memory next to 96GB of GDDR7. If your AMBER or GROMACS jobs are running today on a 24GB or 32GB card, moving to 96GB rarely unlocks a system size you could not already simulate.
The VRAM ceiling in a structural biology pipeline is almost always set by structure prediction, not by the MD engine. That is where the hard numbers live, and DeepMind publishes them.
AlphaFold 3, per DeepMind’s own documentation
- Inputs up to 5,120 tokens are verified to fit on a single 80GB A100 or H100.
- Up to 4,352 tokens fit on a 40GB A100, but only with unified memory enabled and a modified
pair_transition_shard_spec, and at reduced throughput. - Beyond 5,120 tokens, or on smaller cards, unified memory lets GPU memory spill into host RAM. This prevents the out-of-memory error at the cost of running slower.
- The genetic search stage is separately RAM-hungry: DeepMind recommends at least 64GB of system RAM, and more for long targets.
One caveat that vendor guides consistently omit, and that a structural biology lab should know before spending: DeepMind has validated AlphaFold 3 numerical accuracy only on A100 and H100 devices, and documents known unresolved numerical issues on other GPU types. That does not mean other cards fail — many labs run AF3 on RTX-class hardware — but it does mean you should validate outputs against a known reference before trusting production predictions on any unvalidated device. We would rather tell you that up front than sell you a card and let you find out.
Question 2: How are you parallelizing? (And the MIG correction)
This is the question that actually shapes the build, and it splits cleanly in two.
One large trajectory — a single big system you want to push as far as possible in wall-clock time. NAMD’s GPU-resident mode is built for this, combining whole GPUs behind fast peer-to-peer communication to accelerate one run. Favor fewer, complete GPUs.
Many independent runs — free energy calculations, replica exchange, ensemble sampling, parameter sweeps. Here you want isolation between trajectories, not raw single-run speed. This is what MIG is for.
Correcting the record on MIG
Several vendor comparisons currently state that Multi-Instance GPU is available only on the RTX PRO 6000 Blackwell Server Edition, and use that to steer ensemble-heavy labs toward a rackmount server. That is not what NVIDIA documents.
NVIDIA lists Multi-Instance GPU as a feature on the RTX PRO 6000 Blackwell Workstation Edition product page, describing up to four fully isolated instances, each with its own high-bandwidth memory, cache, compute cores, and guaranteed quality of service. MIG capacity profiles across the 96GB RTX PRO 6000 family are documented as up to four 24GB instances, two 48GB instances, or one 96GB instance.
The genuine Server-Edition-only feature is NVIDIA vGPU, which is a different thing. NVIDIA’s vGPU documentation states that the Workstation Edition and Max-Q Workstation Edition do not support vGPU, while the Server Edition does, including MIG-backed vGPU configurations. MIG is hardware partitioning of one card. vGPU is a virtualization software stack for multi-user virtual workstations and VDI. MIG is not a substitute for a validated vGPU deployment.
| MIG | NVIDIA vGPU | |
|---|---|---|
| What it does | Partitions one physical GPU into isolated hardware instances | Virtualizes GPU access across multiple VMs / users |
| Workstation Edition | Documented by NVIDIA | Not supported |
| Max-Q Edition | Documented by NVIDIA | Not supported |
| Server Edition | Yes | Yes, including MIG-backed vGPU |
| Right for | Parallel replicas, isolated ensemble runs on one card | Multi-user VDI, shared virtual workstations |
What this means practically: a single-GPU workstation can host several isolated ensemble trajectories without buying a server. If someone told you otherwise, the recommendation was more expensive than it needed to be. Verify MIG profile availability against your driver and platform before committing — supported profiles depend on driver, firmware, and deployment configuration — but do not accept “you need Server Edition for MIG” as a given.
Question 3: Where is the actual bottleneck?
MD engines do not lean on hardware the same way, and this is where builds most often go wrong. Buying the biggest GPU and pairing it with a modest CPU produces a machine that is slower than its parts suggest.
Per the GROMACS user guide, in the default hybrid scheme the CPU handles bonded forces and PME while non-bonded short-range work is offloaded to the GPU. Domain decomposition and trajectory I/O also sit on CPU cores. That makes a weak CPU a hard ceiling on a fast card. AMBER’s pmemd.cuda sits at the other end: it is largely GPU-resident, so clock speed matters more than core count. NAMD 3 has been GPU-resident since version 3, though setup and communication still touch the CPU.
Stop guessing — read the mdrun log
GROMACS tells you the answer for free. At the end of every run, the cycle and timing summary table includes a Wait GPU NB local counter.
Large “Wait GPU NB local” → CPU finished bonded + PME first, is waiting on the GPU
→ the GPU is your bottleneck
→ the GPU is your bottleneck
Small/absent wait + low ns/day → the CPU side is holding you back
→ more or faster cores, or offload PME to GPU
Run this check before you buy anything. Published core-count recommendations for GROMACS vary widely — offloading PME to the GPU changed the calculus substantially, and the right number depends on your version, system, and mdp settings. Your own log file beats anyone’s rule of thumb, including ours.
| Engine | Where the work sits | What to prioritize |
|---|---|---|
| AMBER (pmemd.cuda) | Largely GPU-resident; low CPU load | GPU capability and clock speed over core count |
| GROMACS | Hybrid — bonded and PME on CPU by default, non-bonded on GPU | Balanced CPU; measure before sizing cores |
| NAMD 3 | GPU-resident, scales across whole GPUs peer-to-peer | Complete GPUs over partitioning; adequate CPU for setup |
| LAMMPS | Varies heavily by package and pair style | Balanced system; profile your specific input |
| AlphaFold 3 | CPU-heavy genetic search, then GPU-heavy inference | Both, plus 64GB+ RAM and fast storage |
The parts nobody budgets for: RAM and storage
Two line items get cut from research builds and cause problems within the first month.
System RAM. DeepMind recommends at least 64GB for AlphaFold 3’s genetic search stage, more for long targets. Separately, every parallel MD trajectory holds its own topology, coordinate, and force arrays in system memory. Four concurrent runs plus an analysis pipeline is a different RAM requirement than one run. Use ECC — a silent bit flip in a week-long trajectory is discovered days after it happened.
Storage. AlphaFold 3’s genetic databases alone can require up to 1TB of disk, with SSD recommended. MD trajectory output compounds fast, and several runs writing simultaneously will saturate a single SATA SSD. NVMe for scratch and active trajectories, bulk capacity for archive.
Sizing the build
| Your workload | Configuration | Build |
|---|---|---|
| Single-user lab, mixed AMBER/GROMACS, moderate systems | 1x RTX PRO 6000 Blackwell, balanced CPU, ECC DDR5 | Threadripper PRO · Intel Xeon |
| Ensemble / replica exchange / free energy | 1-2x RTX PRO 6000 with MIG partitioning | Threadripper PRO |
| Large single trajectories, NAMD multi-GPU | 2-4x RTX PRO 6000 Max-Q, PCIe Gen 5 x16 per card | Threadripper PRO tower · 5U rack |
| Structure prediction pipeline (AF3 + MD) | 96GB GPU, high system RAM, 1TB+ NVMe for databases | Threadripper PRO · EPYC |
| Shared group resource, many users | 4-8x GPU rackmount, vGPU if multi-user virtualization needed | 4U EPYC server |
VRLA Tech at vrlatech.com builds computational chemistry and structural biology workstations to order. Tell us which engines dominate your week, your typical system size, and whether you run one big trajectory or many replicas — an engineer will size CPU, GPU count, RAM, and storage against that rather than against a spec sheet. Most workstations and servers ship within 2-3 weeks, with a 3-year parts warranty and lifetime US-based engineer support. Built in Los Angeles since 2016. Customers include Los Alamos National Laboratory, Johns Hopkins University, General Dynamics, Miami University, and George Washington University.
Ready to buy?
Hardware and specification questions
- Does the RTX PRO 6000 Blackwell Workstation Edition support MIG?
- Yes. NVIDIA lists Multi-Instance GPU on the RTX PRO 6000 Blackwell Workstation Edition product page, supporting up to four fully isolated instances, each with its own memory, cache, and compute cores. Some vendor comparisons incorrectly state MIG is Server Edition only. The genuine Server-only feature is NVIDIA vGPU, which is different from MIG. VRLA Tech configures MIG partitioning on workstation and server builds in Los Angeles. Since 2016, 3-year parts warranty, lifetime US-based engineer support.
- What is the difference between MIG and vGPU?
- MIG partitions one physical GPU into isolated hardware instances with dedicated memory and compute, and is available on the RTX PRO 6000 Workstation, Max-Q, and Server editions. NVIDIA vGPU is a virtualization software stack for VDI and multi-user virtual workstations, and is supported only on the Server Edition. MIG is not a substitute for a validated vGPU environment. VRLA Tech configures both. Built in Los Angeles since 2016 with a 3-year parts warranty and lifetime US-based engineer support.
- How much GPU memory does AlphaFold 3 need?
- DeepMind has verified inputs up to 5,120 tokens on a single 80GB A100 or H100. Up to 4,352 tokens fit on a 40GB A100 with unified memory enabled and a modified pair transition shard spec, at reduced throughput. Larger inputs require unified memory, which spills to host RAM and runs slower. VRLA Tech sizes structure prediction workstations around these limits. Los Angeles since 2016, 3-year parts warranty, lifetime US-based engineer support. Customers include Johns Hopkins University.
- Is VRAM the main constraint for classical molecular dynamics?
- Usually not. Classical MD in AMBER, GROMACS, and NAMD is comparatively modest in GPU memory even at large atom counts, and a 96GB card leaves substantial headroom. The VRAM ceiling in a structural biology pipeline is far more often set by structure prediction tools like AlphaFold 3 than by the MD engine itself. VRLA Tech sizes from the whole pipeline, not one application. Built in Los Angeles since 2016. 3-year parts warranty and lifetime US-based engineer support.
- Do I need ECC memory for molecular dynamics?
- For multi-day production trajectories, yes. A single-bit flip can silently corrupt a run you only discover days later, and restarting a week of compute costs far more than ECC does. The RTX PRO 6000 Blackwell ships with ECC GDDR7, and VRLA Tech pairs it with DDR5 ECC system memory on every research build. Built in Los Angeles since 2016. 3-year parts warranty and lifetime US-based engineer support. Customers include General Dynamics and Los Alamos National Laboratory.
- How much storage does a molecular dynamics lab need?
- More than most builds budget for. AlphaFold 3 genetic databases alone can require up to 1TB of disk, and SSD storage is recommended. MD trajectory output compounds quickly, and several parallel runs writing at once will saturate a single SATA drive. VRLA Tech configures NVMe scratch plus bulk project storage on research workstations. Los Angeles since 2016, 3-year parts warranty, lifetime US-based engineer support.
Engine and buying questions
- How many CPU cores do I need for GROMACS?
- It depends on offload configuration, so measure rather than guess. In the default hybrid scheme GROMACS runs bonded forces and PME on the CPU while non-bonded work goes to the GPU, which means an underpowered CPU throttles a fast card. Published core-count guidance varies widely with version and settings. VRLA Tech builds Threadripper PRO and EPYC platforms and helps size cores to the engine. Los Angeles since 2016, 3-year parts warranty, lifetime US-based engineer support.
- How do I tell if my CPU or GPU is the bottleneck in GROMACS?
- Read the cycle and timing summary table at the end of the mdrun log. A large Wait GPU NB local value means the CPU finished its bonded and PME work first and is waiting on the GPU. Small or absent wait time with low ns/day points at the CPU side instead. This diagnostic costs nothing and tells you which component to upgrade. VRLA Tech sizes builds from measured bottlenecks. Built in Los Angeles since 2016. 3-year parts warranty and lifetime US-based engineer support.
- Should I run one big simulation or many replicas?
- This choice drives the hardware more than any spec. Free energy calculations, replica exchange, and ensemble sampling want many isolated instances, which favors MIG partitioning or multiple GPUs. A single very large trajectory favors fewer, complete GPUs with fast interconnect. VRLA Tech configures both patterns. Los Angeles since 2016, 3-year parts warranty, lifetime US-based engineer support. Customers include Los Alamos National Laboratory.
- What workstation should I buy for GROMACS, AMBER, and NAMD?
- Size the CPU to the engine and the GPU count to your parallelization pattern. A Threadripper PRO platform with one or two RTX PRO 6000 Blackwell cards covers most single-lab workloads, while ensemble-heavy groups benefit from more GPUs or MIG partitioning. VRLA Tech builds these to order. Most systems ship within 2-3 weeks. Built in Los Angeles since 2016. 3-year parts warranty and lifetime US-based engineer support.
- Where can I buy a molecular dynamics workstation?
- VRLA Tech builds molecular dynamics and computational chemistry workstations to order in Los Angeles, sized from the engines you run and how you parallelize rather than from a spec sheet. Every system ships burn-in tested with CUDA and your simulation stack configured. Most workstations and servers ship within 2-3 weeks. 3-year parts warranty and lifetime US-based engineer support. Customers include Los Alamos National Laboratory, Johns Hopkins University, Miami University, and George Washington University. In business since 2016.
For the full RTX PRO 6000 variant breakdown, see Workstation vs Max-Q vs Server Edition. For multi-GPU thermal validation, see our 4x RTX PRO 6000 test results. For GPU memory math on AI workloads, see Agentic AI VRAM Requirements. For research computing at scale, see HPC servers for research labs and the scientific computing workstation guide.
VRLA Tech builds for research laboratories, pharmaceutical and biotech, healthcare, defense and government, and finance organizations.
