Great science. Less machinery.
There’s remarkable science inside bioinformatics tools. We’re rethinking how it runs. Our first application, nomics, takes human germline reads from FASTQ to VCF in one Rust process on CPU and GPU, and checks every stage against nf-core/sarek.
Coming soon: a first look at nomics for MacFaster, without
changing the answer.
nf-core/sarek 3.10.0 and nomics on HG002 and GRCh38, one DGX Spark, October 2026. The best recorded run for each configuration, with dates and memory state kept in view—not an estimate of typical runtime.
HG002 30× whole genome · raw paired FASTQ → VCF
11.34×observed sarek / nomics wall-time ratio
nomics: 41 / 41 stage artifact gates passed per run.
Best run: first after reboot. Three later runs of the same build on fragmented memory took 3,930–4,024 s.
sarek and nomics use matching preparation and reference-gated scientific outputs. Reporting and QC coverage differ; the checks are scoped to tested stage artifacts.
Runs, configuration and comparison limits
Observations behind the bars
| Configuration | Observed runs |
|---|---|
| sarek | 41,044 |
| nomics, fused | 3,619; later 3: 3,930–4,024 |
sarek has one observation. The nomics best is the first run after reboot, followed by three same-build runs reported as a 3,930–4,024 s range. This is not an estimate of typical runtime.
Same machine; different campaign dates
sarek ran 17–18 October and this nomics build ran on 27 October. The original sarek run predates the Linux update; these are dated results, not one kernel-controlled campaign.
Runs used the same DGX Spark, one job at a time, with page caches evicted. The best 27 October nomics run was first after reboot; later runs on fragmented memory took longer. Build, date and memory-state differences matter.
Which reference, and which checks?
sarek 3.10.0 on arm64; the matching x86 configuration uses Java PairHMM and Smith-Waterman with HotSpot libm intrinsics disabled. Default x86 Intel GKL differs on about 2% of 30× records. Reference variability.
Checks apply to tested scientific stage artifacts: bytes, decoded records, or explicitly normalized provenance. Compression and some headers differ; QC/report coverage is not equivalent. This is not a claim that every file is byte-identical. CRAM output now has passing 30× decoded-record, header, index and region checks. The plotted historical runs use BAM-mode output; CRAM codec bytes are not required to match.
Read the results and measurement history, reproduction guide, and artifact comparison rules. Exome and comparator observations come from the 17–18 / 23 October 2026 campaign; the best genome run is from 27 October 2026. Values are reported seconds. The exome plot selects 290 s, not the 298 s mean used in the README.
A pipeline is more
than its plumbing.
The algorithms matter. So do all the things between them.
Workflow engines make complex analyses possible. But for a well-defined workflow, how much orchestration could become one coherent application? How much data could stay close to the next stage?
Baremetal Bioinformatics explores those questions from the implementation up. We build on established scientific functionality, reimplement execution, and measure concordance against pinned references. Changes earn their place through evidence.
The same analysis.
Fewer boundaries.
A workflow is a network of processes, files and synchronization points. Here’s the measured germline path—not just a row of algorithms.
nf-core/sarek 3.10.0 · measured WGS configuration
R1 / R2 · FASTQ.gz12 × paired FASTQ.gz → read in new taskschunk.bamchunk.bamchunk.bampublished sorted BAMmd.bam + index + metricstables × groupsrecal.tableBAMs × groupsrecal.bamVCFs × groupssample.vcf.gz + indexReference FASTA, BWA index, known sites and interval files feed the relevant tasks. QC/reporting is another branch: FastQC, metrics and MultiQC are omitted from this core-path drawing.
nomics one purpose-built Rust process
Bounded batches overlap on shared workers and CPU/GPU-accessible buffers.
Keep chunk, batch and interval semantics. Change the scheduling and data movement.
Reference checks alongside
The validation harness compares stage artifacts with the pinned sarek run, using explicit comparison rules. Differences stay documented.
Reference tools are test oracles—not runtime subprocesses inside nomics.The measured 30× sarek trace had 116 tasks, including support and reporting work. The drawing groups related tasks; it is not a one-box-per-task trace.
nomics
Flagship project · in developmentHuman germline analysis, from paired FASTQ reads to variant calls, in one purpose-built Rust application.
A concrete place to test the thesis: connect the stages, reduce repeated work, and check the results against sarek 3.10.0 on arm64 and its matching x86 Java HaplotypeCaller configuration.
- Prepare readsfastp semantics
- Align + sortBWA-MEM semantics
- Mark duplicatesGATK MarkDuplicates
- Recalibrate base qualitiesGATK BQSR
- Call variantsGATK HaplotypeCaller
A little omics powerhouse.
Closer to home.
nomics for macOS / Metal Coming soon · sneak peek
An early look at full-scale runs on Apple Silicon, with Metal and the same reference checks. Still experimental; not a generally supported release.
Read the early Metal notes- Metal prototype
- 1h 16m 32s 4,592 s
- CPU control, same Mac
- 2h 49m 21s 10,161 s
2.21× observed CPU / Metal wall-time ratio
Both runs passed 41 / 41 normalized artifact checks against the pinned reference.
One matched 30× run per configuration, with tuned CPU/GPU work sharing. Early research evidence, not a supported release.
Working code.
Open questions.
Evidence belongs next to scope. nomics is research software you can build and inspect; demonstrated output agreement is not a claim of clinical or production validation.
Working on tested fixtures
- Single-sample germline analysisWES and WGS through VCF/GVCF, with BAM or CRAM output. Exome single-sample joint genotyping is also gated.
- CPU + CUDA execution on DGX SparkThe primary measured platform. Large genomes spill and stream through disk.
- Multi-lane and CPU-only coverageOptical-duplicate fixtures and a gated 30× x86 CPU run, alongside the primary arm64 reference.
- Documented build and reproduction pathsA source build, pinned references, stage gates and digest manifests—not a packaged, generally supported product.
Coming next and experimental
- macOS / Metal: coming soonA sneak peek backed by early full-scale M4 Max results. The broader Mac experience is still being developed.
- RNA-seq and single-cell componentsSalmon alignment mode, featureCounts and STARsolo oracle-BAM counting are experimental. No complete STAR aligner; not yet faster than their references.
- Integration with existing pipelinesnf-core shim task replays exist; complete workflow integration is still in progress.
- Reusable pieces for your own workflowBuild on formats, stage implementations, shared-memory execution and reference checks. Public APIs are still developing.
Tested on HG002 and the iGenomes GATK GRCh38 bundle with pinned tool versions. Somatic calling, other references and other tool versions are not supported by the current compatibility claim.
Questions,
answered plainly.
- What is nomics?
- A from-scratch Rust implementation of the human germline short-read workflow: read preparation, BWA-MEM alignment and sort, duplicate marking, base quality recalibration and HaplotypeCaller variant calling, in one process on CPU and NVIDIA CUDA GPUs. Its outputs are checked stage by stage against nf-core/sarek 3.10.0.
- How much faster is it than nf-core/sarek?
- On one DGX Spark, the best measured HG002 30× genome run took 1 h 00 m from paired FASTQ to VCF, against 11 h 24 m for sarek: an observed ratio of 11.34×. Three later runs of the same build took 3,930–4,024 s. These are dated best observations, not typical-runtime estimates.
- Does it produce the same results?
- On the tested fixtures, yes: 41 / 41 stage artifact gates pass for the 30× genome and 18 / 18 for the exome. Gates compare bytes, decoded records or explicitly normalized provenance, as written in GATES.md. The match is to sarek on arm64 and its matching x86 Java HaplotypeCaller configuration; default x86 Intel GKL differs on about 2% of 30× records.
- Is it clinically validated?
- No. nomics is research software. Agreement with a pinned reference on HG002 and GRCh38 is not clinical or production validation.
- What hardware does it run on?
- The primary measured platform is an NVIDIA DGX Spark with CPU and CUDA execution. A 30× x86 CPU-only run is also gated. macOS / Metal is an experimental sneak peek.
- How do I get it?
- Build it from source. It is licensed GPL-3.0-only; there is no packaged release yet. Questions and bug reports go to GitHub issues.
One application now.
More possibilities next.
nomics is the first application, and a place to build useful foundations: formats, stage implementations, shared-memory execution, and reference checks.
The longer-term ambition is to make those pieces useful beyond one pipeline—so you can rethink the machinery around your own science.
Independent research software · GPL-3.0-only.
Build from source, inspect the evidence, and explore the implementation.