The Ohio State University needed enough compute and data-engineering capacity to assemble a global biodiversity dataset and train a foundation model that could identify species across nearly a million taxa, which was impractical on smaller research infrastructure.
Reported impact
BioCLIP 2 delivered best or top-two species-identification performance across almost one million taxa, improved average species-classification accuracy by 18 percent over BioCLIP, and gave conservation researchers a scalable foundation model trained on the largest biodiversity dataset assembled to date.
Ohio State researchers built BioCLIP 2 around a pipeline that starts with TreeOfLife-200M, a 214 million-image biodiversity corpus assembled from sources including GBIF, Encyclopedia of Life, BIOSCAN-5M, and FathomNet. TaxonoPy standardizes noisy taxonomic hierarchies and distributed download tools place the curated data on HPC file systems, where GPU clusters can keep training fed at scale. The team then trains and evaluates BioCLIP 2 across a mix of NVIDIA A100 and H100 systems, from NAIRR allocations at the Ohio and Pittsburgh supercomputing centers to a 32-H100 primary run for the final model. That architecture gives the researchers a repeatable way to ingest massive ecological datasets, learn species representations across almost one million taxa, and then run recognition and zero-shot identification workloads that outperform prior biodiversity models.
Images and taxonomy data from major biodiversity repositories are normalized with TaxonoPy and stored in TreeOfLife-200M. Training jobs then stream the curated corpus onto A100 and H100 GPU clusters, which produce the BioCLIP 2 model used for large-scale species classification and zero-shot ecological recognition.
18.0% average improvement in species classification accuracy over BioCLIP
As part of the US National Science Foundation National AI Research Resource (NAIRR) pilot program, the team used a combined allocation at the Ohio Supercomputer Center and the Pittsburgh Supercomputing Center, running 64 NVIDIA A100 GPUs for 290 hours.
The primary training run for the final BioCLIP 2 model was conducted on 32 NVIDIA H100 GPUs for 10 days.
Trained on the world's largest biology dataset, comprising 214 million images, using NVIDIA GPUs.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Oct 6, 2026 University of Pennsylvania Penn needed shared AI infrastructure that could support researchers across many disciplines without weeks-long wait times, fragmented storage, or the cost barriers that kept frontier-scale training and inference off… 1.5 years of compute time reduced to three days for a 40,000-document project NVIDIA DGX B200 systems · NVIDIA DGX SuperPOD · NVIDIA InfiniBand NVIDIA Sample Aug 27, 2026 Texas A&M University Texas A&M needed to give researchers across seven institutions access to shared GPU infrastructure for AI and simulation, but isolated workstations, slow grant-funded clusters, and long queues were limiting… $1M saved on a single project DDN high-performance storage · NVIDIA Base Command Manager · NVIDIA DGX SuperPOD NVIDIA Sample Aug 3, 2026 StudyFetch StudyFetch needed to keep AI tutoring affordable while scaling live lecture transcription and personalized learning for millions of students, but off-the-shelf cloud transcription created cost and latency ceilings that… 100M+ learning interactions NVIDIA H100 GPUs · NVIDIA L40S GPUs · NVIDIA NIM 