What each task asks for
Alignment (BWA, Bowtie2, STAR, minimap2) scales with cores and needs tens of gigabytes of memory for the indexes. Variant calling (GATK, DeepVariant) mixes cores and, in DeepVariant, GPU. De novo assembly is the extreme case: hundreds of gigabytes of memory in one machine. Single-cell analysis (Seurat, Scanpy) lives in memory. And basecalling from nanopore sequencers and protein structure models run on the GPU.
Memory: the part that decides
For alignment and variant calling in human, 64 to 128 GB. For single-cell with hundreds of thousands of cells, 128 to 256 GB. For de novo assembly of large genomes or metagenomics, 512 GB or more, on a platform with eight channels (Threadripper PRO) or twelve (EPYC). And always ECC: a memory error in a two-day assembly is a bad result nobody detects.
Cores and GPU
A 32-core Threadripper aligns a whole human genome in a fraction of the time of a desktop machine, and a 64 or 96-core Threadripper PRO serves several pipelines at once. The GPU comes in with nanopore basecalling (CUDA, the more memory the better), with DeepVariant and with deep learning models: an RTX 5090 or a 48 GB RTX PRO 5000 for the large models.
Drives: terabytes of FASTQ and intermediate results
A sequencing run is hundreds of gigabytes of FASTQ, and the intermediates (BAM, VCF) take as much again. A 4 TB NVMe as the working drive, another for the system, and 16 TB mirrored drives or a file server for the archive and backups. And 10 GbE networking if the data comes from the sequencer or a server.
| Work | Workstation | Price |
|---|---|---|
| Alignment, variants and single-cell | Advanced Plus: 32-core Threadripper, 128 GB registered ECC, RTX 5080, 2 TB + 4 TB NVMe | €14,265 |
| Several pipelines at once, basecalling and models | Enterprise: 32-core Threadripper PRO, 256 GB ECC in eight channels, RTX 5090, 2 × 4 TB NVMe + 16 TB | €30,435 |
| De novo assembly and metagenomics | Enterprise Supreme: 64-core Threadripper PRO, 512 GB ECC, 2 × RTX 5090, 3 × 4 TB NVMe + 16 TB | €57,400 |
Linux, containers and the shared machine
Almost everything runs on Linux, with Conda, Docker or Singularity; we deliver the machine with the distribution you use and the GPU drivers installed. If the laboratory shares the machine, a simple queue manager (Slurm on one node) stops two pipelines from trampling each other’s memory. And if the load grows, the same work moves to a rack-mounted EPYC server with more memory and cores.
Related hardware and guides
Frequently asked questions
Workstation or server for genomics?
If one or two people launch the pipelines and want to see the results at their desk, a workstation. If the laboratory shares the load or assemblies last for days, a rack-mounted EPYC server with more memory and cores, shared by everyone.
How much memory for a de novo assembly?
It depends on the genome and the assembler: from 128 GB for small genomes to more than 512 GB for large genomes or metagenomes. Check the peak memory of your assembler with similar data and double it.
Is a GPU needed?
For aligning and calling variants with the classic tools, no. For nanopore basecalling, DeepVariant, protein structure models and deep learning, yes, with as much memory as possible.
Do you install Linux and the tools?
Yes. We deliver the machine with the Linux distribution you use, the GPU drivers, Conda or Docker and the GPU tested for 24 hours at full load. You install the pipelines yourselves or we do it together.
Shall we size your workstation?
Tell us which software you use and the size of your projects. We will propose the processor, memory, GPU and drives, with a price, within 12 working hours.