Cryo-EM GPU server and workstation: cryoSPARC and RELION hardware requirements and sizing

What a cryo-EM processing machine needs

A cryo-EM GPU server or workstation runs the processing after the microscope in cryoSPARC or RELION, which both do the heavy work on NVIDIA GPUs in single precision. cryoSPARC's hardware guide recommends, per GPU, four CPU cores, 64 GB of RAM and a fast SSD for the particle cache, and it states that the memory of each GPU limits the largest box size a reconstruction can use. A workstation with four NVIDIA RTX PRO 5000 Blackwell 72 GB cards, or a rack server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards, meets those rules with room for large boxes.

The processing runs in steps: motion correction and CTF estimation of the movies, particle picking, 2D classification, and 3D refinement and classification. For a director of a structural biology institute or a pharma research unit, the sizing comes down to three figures: the datasets the microscopes produce per week, the particle and box sizes, and the number of people who process data at the same time. One microscope and a small team usually work well with one four-card workstation next to the facility storage. A facility with several microscopes or many users is better served by rack-mounted GPU workers in a server room, shared through cryoSPARC's master and worker model or a scheduler. Our guide to the university research lab server covers scheduling and quotas, and our guide on GPU servers for pharma and life sciences covers structure prediction, which often runs next to cryo-EM in the same department.

cryoSPARC hardware requirements, per GPU worker

We read cryoSPARC's hardware and system requirements page, which covers version 5.0 and later, on 24 October 2026. cryoSPARC separates a master node, which runs the database and web interface and needs no GPU, from worker nodes that run the jobs.

cryoSPARC guide, hardware and system requirements (v5.0+), read on 24 October 2026. The master node is listed with 8 or more cores at 2.8 GHz or above, 32 GB of RAM and a 500 GB SSD recommended, and no GPU.

The GPU row on that page still names compute capability 3.5 as the minimum. The installation prerequisites page is more precise for current versions: cryoSPARC 5.0 and later supports "only GPUs with compute capability 5.0 (Maxwell) to 12.0 (Blackwell)", bundles CUDA 12.8 so that no system CUDA is needed, and requires NVIDIA driver 570.26 or newer. It adds that "NVIDIA Blackwell devices are only compatible with the open driver", which matters for every RTX PRO Blackwell card. All cards in this article are within that range: the RTX PRO 5000, RTX PRO 6000 and RTX PRO 4500 Blackwell at 12.0, the L40S at 8.9 and the H200 NVL at 9.0.

The memory rule is the one that decides the card. cryoSPARC states: "The GPU memory (VRAM) in each GPU limits the maximum particle box size for reconstruction", and "Typically, a GPU with 12GB VRAM can handle a box size of up to 700^3, and up to 1024^3 in some job types." It gives no figure for larger cards. A 3D map grows with the cube of its box edge: in 32-bit floats a 700³ map holds 1.4 GB and a 1024³ map 4.3 GB by our arithmetic, and jobs keep several such volumes and their Fourier transforms at once. Labs that work on large complexes, viruses or filaments at fine pixel sizes are the ones that run into the limit, and for them 72 or 96 GB per card is the safe choice.

RELION 5 on Blackwell cards

RELION is open source under the GPL 2.0 licence, and its installation page for release 5.0 states that it builds with GPU acceleration "by default", provided CUDA is found. As of RELION 5.0.1 the developers "test RELION builds with CUDA 11 to 13". Two details matter on new hardware. The build option CUDA_ARCH defaults to 35, compute capability 3.5, and the page warns that with a wrong number "compilation might succeed but the resulting binary can fail at the runtime"; RTX PRO Blackwell cards have compute capability 12.0, which is 120 in the same notation. And the Python tools Blush, ModelAngelo and DynaMight run in a conda environment, for which RELION ships environment_blackwell.yml for "NVIDIA GPUs of the Blackwell generation (50x0) or newer".

The precision question is settled by RELION itself. A double-precision GPU build "will slow down GPU-execution considerably" and does not improve resolution, according to the installation page, so it is off by default. Cryo-EM processing therefore has no use for the fast FP64 of an H200 NVL; our article on FP64 GPU servers explains which workloads do, and our guide to GPU servers for chemistry and materials research covers the molecular dynamics and DFT codes that often run in the same institute.

RELION's processes and memory follow its tutorial. 3D auto-refine needs at least three MPI processes and runs best with an odd number, because one process coordinates the others; the tutorial's example on "our computer with 4 GPUs" used five MPI processes with six threads each and took about seven minutes on the small tutorial dataset. Padding of 3D maps, on by default, "requires 8x more computer memory (RAM)", and memory use rises again in the final iteration. The tutorial's layout of five processes with six threads each uses 30 threads, so a four-card RELION machine needs more cores than cryoSPARC's four per GPU, and generous host memory; our configurations below provide both.

Tell us your typical box size and particle count, with the programs your team uses, and our team replies with a card and memory configuration that fits them.

Card choice: RTX PRO 5000, RTX PRO 6000 or RTX PRO 4500

NVIDIA product pages and datasheets as summarised in our card guides, read 24 October 2026. The placement column is our recommendation for cryo-EM machines.

The RTX PRO 5000 72 GB is the card we recommend for a cryo-EM workstation. It has six times the 12 GB that cryoSPARC quotes for a 700³ box, draws 300 W, and four of them fit one tower on one office circuit. The 48 GB version suits labs whose boxes stay moderate, and our comparison of the RTX PRO 5000 48 GB and 72 GB sets out the difference. The RTX PRO 6000 Max-Q gives 96 GB per card at the same 300 W and is the step up for the largest boxes in a workstation.

For a rack, the RTX PRO 6000 Server Edition is the card to choose: it is passive, built for server airflow and continuous load, and gives 96 GB. The RTX PRO 5000 is an active workstation card and belongs in a rack only in a chassis whose maker lists it, as our guide to RTX PRO 5000 rack servers explains. The RTX PRO 4500 Server Edition, with 32 GB at 165 W, suits pre-processing workers: cryoSPARC's guide notes that nodes doing only motion correction, picking and CTF estimation need no SSD cache, and pre-processing does not reconstruct the 3D maps whose box size cryoSPARC ties to GPU memory.

Storage: SSD cache, project space and the microscope link

cryoSPARC caches particle stacks on a local SSD on each worker. Its SSD caching guide gives the formula: dataset size equals bytes per pixel times box size squared times the number of particles, plus file headers. Its own examples: one million float16 particles at box 256 take 131 GB, two million float32 particles at box 432 take 1.5 TB. The guide recommends "2TB capacity SSDs, or larger" for the largest stacks, and re-uses a cached stack when several jobs in a project need the same particles. We fit four-card workers with at least 4 TB of NVMe for the cache, so that two large stacks fit side by side.

The project data lives on shared storage that the master and all workers can reach. For pre-processing, cryoSPARC gives a rule of thumb: "to saturate a 4-GPU machine during pre-processing, a sustained sequential read of 1000MB/s is required." A 10 Gbps link carries 1,250 MB/s by our arithmetic, so a four-card worker reading movies from a file server needs at least a 10 Gbps connection and storage that sustains that rate; 25 Gbps leaves room for several workers. Our guide to NVMe storage for a GPU server covers drive classes and endurance for cache drives that are rewritten daily.

Operating system, licences and network

cryoSPARC 5.0 and later needs GLIBC 2.28 or newer and runs on x86-64 Linux. The guide recommends Ubuntu 22.04 or 24.04 or a current LTS release such as Rocky Linux 8, 9 or 10, and warns: "As of June 2026, CryoSPARC versions up to 5.0.6 are incompatible with the version 7 kernel", the default of Ubuntu 26.04. We therefore deliver cryo-EM machines with Ubuntu 24.04 LTS or Rocky Linux and the open NVIDIA driver, unless your facility standard says otherwise.

Licences differ between the two programs. cryoSPARC is "available free of charge for non-profit academic use"; companies, contract research organisations and pharma companies request a commercial licence through the cryoSPARC download page. A licence ID is entered at installation, the worker's ID must match the master's, and the software needs outbound HTTPS access to the licence server, which a closed network has to allow. RELION is open source under GPL 2.0. cryoSPARC's guide also states that it "is designed to be run only within a trusted private network" and should never be hosted directly on the internet or an untrusted network "without a separate controlled authentication layer"; our guide to securing a GPU server covers the network side.

Three configurations for a cryo-EM lab

Our configurations; host memory follows NVIDIA's configuration guide for certified systems (twice the GPU memory), above cryoSPARC's 64 GB per GPU. Wall power is our estimate as in our four-card workstation and server guides: cards plus processors plus 15 per cent; a measurement on the delivered system confirms it.

By cryoSPARC's rule a four-card worker needs 16 cores and 256 GB of RAM. Our hosts give more: the 32-core Threadripper PRO and two 32-core EPYC 9355 cover RELION's five MPI processes with several threads each, and 768 GB leaves room for padded RELION maps and large classifications. The four-card workstation with RTX PRO 5000 72 GB is described in detail in our guide to four RTX PRO 5000 72 GB in one workstation; it draws about 1.8 to 2 kW at the wall, under 9 A on 230 V. A rack worker with four RTX PRO 6000 Server Edition cards at 600 W needs about 3.4 kW, more than one 230 V 16 A feed planned at 2.94 kW, so it goes on a 32 A or three-phase feed. Several rack workers can share one master on a small server or virtual machine.

How we deliver the cryo-EM machine

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 workstation or 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, the NVIDIA driver, CUDA and a container runtime, and commission the machine on site. cryoSPARC is installed with your own licence ID. Quantities and lead time are confirmed in writing with the quote.

To quote, we need the microscopes and detectors feeding the machine, datasets per week, typical and largest box sizes, the programs and versions you use, the number of people processing at the same time, the facility storage and network, and whether the machine stands in an office or a server room. Send us these details, and we reply with a configuration and a written quote. Labs that also run genome analysis will find the matching sizing in our guide on the GPU server for genomics with Parabricks.

FAQ