Faster and cheaper at the same time.
Field-programmable gate arrays suit the specific computational shape of read alignment and variant calling. Because F2 instances are priced comparably to the CPU instances they replace, the speedup converts directly into a lower cost per genome.
01 · Why FPGAs
The work has the right shape.
Alignment and variant calling are dominated by highly parallel, fixed-pattern operations over enormous streams of short reads. That is precisely the workload a reconfigurable hardware fabric executes well, and precisely where a general-purpose CPU spends most of its transistor budget on things the problem does not need.
AWS cites up to 30× NGS compute acceleration for FPGAs versus current CPU instances. That is AWS's figure for FPGA acceleration in general, not a GenoAccel measurement — an important distinction we keep explicit.
Counting input/output and the pipeline stages that do not accelerate, GenoAccel targets a conservative 3–5× end-to-end improvement on a full whole-genome pipeline, with a design goal of completing a 30× whole human genome in under six hours against a CPU baseline of roughly sixteen.
02 · Why F2, and why now
The previous generation is gone.
AWS F1 instances reached end-of-life on 31 December 2025, and AWS migrated customers to F2. This is not a minor version bump: every cost model, instance sizing and bitstream path built on F1 is void.
F2 is the second generation — up to eight AMD Virtex UltraScale+ HBM VU47P FPGAs, 16 GB of high-bandwidth memory, third-generation AMD EPYC hosts and 100 Gbps networking — and AWS positions it at materially better price-performance than F1.
Critically, the commercially maintained bitstream path runs on F2: Illumina deploys DRAGEN secondary-analysis pipelines on F2 through Illumina Connected Analytics. That establishes the hardware path as viable rather than experimental.
03 · Our own numbers
Not published until measured.
A genomics audience checks claims. Publishing a plausible number now would cost more credibility than the number could ever buy.
| Metric | Value | Sample | Build | Reference set | Pipeline | State |
|---|---|---|---|---|---|---|
| 30× WGS wall clock | — | NA12878 | GRCh38 | — | — | Pending |
| SNV F1 score | — | NA12878 | GRCh38 | — | — | Pending |
| Indel F1 score | — | NA12878 | GRCh38 | — | — | Pending |
| Cost per genome | — | — | — | — | — | Pending |
The distinction that matters. “Up to 30×” is AWS's general FPGA figure. “Under six hours” is our design target. Neither is a GenoAccel measurement, and we will not present either as one. When the benchmark suite runs, each figure will be published with its sample, genome build, reference set and pipeline version, alongside a reproducible methodology.