GPU server for genomics: NVIDIA Parabricks requirements, genomes per day and storage

What a genomics GPU server does

A GPU server for genomics runs secondary analysis, from aligning a sequencer's reads to calling variants, and the documentation of NVIDIA Parabricks 4.8.0 sets its hardware requirements: GPUs with 16 GB or more each, two or more per server, at least 100 GB of host RAM and 24 CPU threads for two GPUs, and fast local SSDs. On four NVIDIA RTX PRO 6000 Blackwell Server Edition cards NVIDIA states that germline analysis drops "from ~16 hours to under 10 minutes".

Secondary analysis aligns the reads from a sequencer to a reference genome, sorts them, marks duplicates and calls variants. Parabricks runs these steps on GPUs; its documentation calls one GPU "supported but not recommended".

For a director of a sequencing lab, biobank or clinical genetics unit, the question is how many genomes per day the lab must turn around and how much of the work stays in-house. Each Parabricks run uses a whole server, or a defined set of its GPUs, for a few minutes per genome. More cards shorten each run, more servers multiply the number of genomes a day, and the storage in front of the cards decides whether the GPUs wait for data. This guide sizes the server from NVIDIA's published requirements and timings. Structure prediction and drug discovery models are covered in our guide on GPU servers for pharma and life sciences, and a shared research server with scheduling and quotas in our guide for a university research lab.

Parabricks 4.8.0 hardware requirements

We read NVIDIA's installation requirements for Parabricks 4.8.0 on 24 October 2026. The GPU must have "at least 16 GB of GPU memory" and one of the compute capabilities the page lists, which include 8.9 (L4, L40S), 9.0 (H200) and 12.0 (RTX PRO Blackwell). Host memory and threads scale with the number of GPUs.

NVIDIA Parabricks 4.8.0 installation requirements, read on 24 October 2026. The page adds that giraffe and pangenome_germline may need more RAM and threads because of large pangenome index files.

Some tools need more than the 16 GB minimum in their default settings. HaplotypeCaller and MutectCaller need at least 18 GB per GPU, minimap2 also 18 GB, and rna_fq2bam 24 GB; each has a low-memory flag for smaller cards. The RTX PRO 6000 and RTX PRO 4500 cards clear all of these defaults. The L4 has exactly 24 GB, the rna_fq2bam default; NVIDIA's table includes an rna_fq2bam run on L4 but does not say whether it used the low-memory flag.

The software side is short. Parabricks ships as a Docker image, "freely available as a public container on NGC", and needs Docker 20.10 or newer, a Linux distribution that supports the NVIDIA Container Toolkit, and a driver compatible with CUDA 13.3.x. Enterprise support with defined response times is sold through NVIDIA AI Enterprise, which we supply as licences on our NVIDIA AI Enterprise page; the H200 NVL includes a five-year subscription.

Two points change how the cards are shared. The requirements page states that time-sliced vGPU is supported and Multi-Instance GPU is not, so a genomics pipeline gets whole cards, never MIG slices. And a single-GPU system still has to meet the two-GPU host requirements, which tells you NVIDIA expects at least two cards per job.

Which cards NVIDIA tested, and how fast they are

NVIDIA's tested list for 4.8.0 runs from the T4 to the Rubin GPUs and includes four cards we supply or build into servers: the RTX PRO 6000 Blackwell Server Edition, the RTX PRO 4500 Blackwell, the L4 and the H200. The list names the RTX PRO 4500 without an edition; we build servers with the RTX PRO 4500 Server Edition, the passive 165 W card with 32 GB. It names the H200 without saying whether the NVL or the SXM version was tested, and it lists the L40 rather than the L40S.

NVIDIA's Parabricks overview page publishes timings for 4.8.0 on two and four cards of three types. They come from NVIDIA's Perflab on internal nodes, and NVIDIA calls them "for reference only", since speed depends on the dataset, host processor and memory.

NVIDIA Parabricks overview page, benchmark table for v4.8.0 (NVIDIA Perflab), read on 24 October 2026. DeepVariant: 30× Illumina whole genome; minimap2: 35× PacBio whole genome; fq2bam_meth: ENCODE bisulfite data; rna_fq2bam: a melanoma sample from an Illumina NovaSeq 6000. The page names no dataset for the fq2bam row. Its table labels the cards "RTX PRO 6000" and "RTX PRO 4500"; its feature list names the RTX PRO 6000 Server Edition. Two-card results for RTX PRO 4500 and L4 are on the same page.

The table shows that the cards scale unevenly by step. Doubling from two to four RTX PRO 6000 cards cuts DeepVariant by 40 per cent and the bisulfite pipeline by 42 per cent, but fq2bam by only 19 per cent and RNA alignment hardly at all. Four RTX PRO 4500 cards come close to four RTX PRO 6000 on DeepVariant and RNA work, and fall behind on the alignment-heavy rows. The L4, at 72 W per card, takes about three times as long as the RTX PRO 6000 for fq2bam; NVIDIA labels it "Lowest Cost Per Sample".

Genomes per day: what the timings mean

A whole-genome germline run with DeepVariant is fq2bam followed by DeepVariant. Adding the two rows gives the GPU time per genome, and dividing a day by it gives an upper bound for one server. That bound ignores copying FASTQ files in, quality control, joint genotyping, idle time between jobs and the fact that NVIDIA's fq2bam dataset is not named.

Our arithmetic from NVIDIA's v4.8.0 table: fq2bam plus DeepVariant minutes, 1,440 minutes a day divided by the sum. An upper bound, not a measurement.

Two conclusions follow. The first is that throughput grows faster with servers than with cards: two servers with two RTX PRO 6000 Server Edition cards each reach about 214 genomes a day by this arithmetic, one server with four about 150. A four-card server can also run two jobs of two GPUs each, with --num-gpus or NVIDIA_VISIBLE_DEVICES assigning the cards, but then each job needs its own share of host RAM, threads and disk bandwidth, and NVIDIA publishes no figure for that layout. Measure it on your data before relying on it.

The second is that a lab running a few hundred genomes a month has no need for eight cards. Sequencing runs arrive in batches, and a two-card server at 13.44 minutes per genome clears a batch of 40 genomes in about nine hours by the same arithmetic. A four-card server is the choice when turnaround per sample matters, for instance in a clinical setting where one urgent genome should finish in under ten minutes. We build both; send us your genomes per week and the pipelines you run, and our team replies with a configuration sized to that volume.

Processors, memory and storage around the cards

Parabricks puts more load on the processors than the minimum thread counts suggest. NVIDIA's performance guide states that fq2bam starts 16 primary CPU threads per GPU by default for BWA-MEM, so four GPUs use 64 threads for that step alone, twice the 32-thread minimum. We therefore pair four cards with two 32-core AMD EPYC 9355, 128 threads in all, and two cards with smaller EPYC processors. The guide also recommends the performance governor and a fixed maximum CPU frequency, and notes that these settings do not survive a reboot. Our guide to EPYC or Xeon for a GPU server explains the platform choice.

Host memory is less of a constraint than it looks. Our reference configurations follow NVIDIA's configuration guide for certified systems, at least twice the GPU memory: 384 GB for two RTX PRO 6000 Server Edition cards and 768 GB for four. That is four times the Parabricks minimum and leaves room for giraffe and the pangenome pipeline, which NVIDIA says may need more.

Storage decides whether the cards stay busy. NVIDIA's performance guide says "the best performance is achieved when both kinds of files", inputs and outputs as well as temporary files, "are on a fast, local SSD". If inputs live on network storage, temporary files should still go to a local SSD with --tmp-dir. NVIDIA's tests showed good performance with Lustre for input and output files with up to four GPUs; above four GPUs it highly recommends local SSDs for all files. NVIDIA's whole-genome tutorial with the HG002 sample at 30× asks for 512 GB of disk for one walkthrough, reference and indexes included, so a server that processes several genomes in a row needs local NVMe in the terabytes, with finished BAM or CRAM files and VCFs moved to the lab's storage. Our guide to NVMe storage for a GPU server covers drive classes and endurance.

Four configurations for a sequencing lab

These start from the reference configurations on our AI servers page, adapted to the Parabricks rules above.

Reference configurations of our AI servers page and NVIDIA's configuration guide for certified systems (2× GPU memory, at least six cores per GPU). Wall power is our estimate: cards plus processors plus 15 per cent for the rest, as in our sibling guides; a measurement on the delivered system confirms it.

The RTX PRO 6000 Server Edition is the default, because NVIDIA's table shows it fastest on every step and its 96 GB covers every tool's default settings. The RTX PRO 4500 Server Edition suits a lab that mainly runs DeepVariant and RNA alignment and has a small power feed; the 256 GB of host RAM in our four-card build is above the 196 GB that Parabricks asks for. Four L4 cards stay within one small feed; their 256 GB of host RAM clears the 196 GB minimum for four GPUs, and their 24 GB per card sits exactly at the rna_fq2bam default. The H200 NVL appears in NVIDIA's tested list as the H200 but not in its timing table. We propose it only when the same server also trains genomic language models, a workload our pharma guide covers.

Power and cooling follow the cards. Two RTX PRO 6000 Server Edition cards fit a 230 V 16 A feed, which we plan at 80 per cent, 2.94 kW; four need a 32 A or three-phase feed, as our guide to GPU server power on 16 A, 32 A and three-phase explains. A genomics server runs at full load while a batch is processed, so the room and feed are planned for that peak, not for the average.

Genetic data, GDPR and the EHDS

Article 4(13) GDPR defines genetic data as "personal data relating to the inherited or acquired genetic characteristics of a natural person", and Article 9(1) lists genetic data among the special categories whose processing is prohibited unless an exception in Article 9(2) applies, such as scientific research under Article 9(2)(j) with suitable safeguards. Article 9(4) lets member states "maintain or introduce further conditions, including limitations" for genetic, biometric and health data, so national rules can add requirements. Article 35(3)(b) requires a data protection impact assessment for processing "on a large scale" of special categories. Which basis applies to a lab, and what national law adds, is for its data protection officer and legal counsel to confirm; this section is general information as of October 2026.

The European Health Data Space Regulation, (EU) 2025/327, entered into force on 26 March 2025. The Commission's page states that its rules on secondary use start to apply for most data categories in March 2029 and for the remaining ones, such as genomic data, in March 2031. On the research side, the Commission's 1+ Million Genomes initiative "aims to enable secure access to genomics and the corresponding clinical data across Europe"; the Commission states that 25 EU countries, the UK and Norway have signed its declaration, and that the database of its Genome of Europe project will include data of at least 100,000 citizens.

For the server, the consequence is simple: the FASTQ files, alignments and variant calls stay on hardware the institution controls, in its own data centre or a hosting room it chooses. Parabricks ships as a container image, which can be mirrored into a local registry when the server sits in a closed network; our guide to air-gapped installation shows how images and drivers reach such a network.

How we deliver the genomics 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.

To quote, we need the genomes or exomes per week, the read type (short or long reads, bisulfite, RNA), the Parabricks tools you run, the turnaround you need per sample, where FASTQ files arrive from, and the power feed at the rack position. Write to us with these figures, and we reply with a configuration and a written quote. Before production use, our guide to GPU server acceptance testing describes the checks your team can run on delivery.

FAQ