NVIDIA unveiled a new set of CUDA-X libraries and microservices at the ISC conference in Hamburg that target scientific workloads, with the headline result a 14,900x acceleration in loading and reading FITS image data from the Rubin Observatory's Legacy Survey of Space and Time. The same cuPhoton reference code delivered up to 8,400x faster signal processing and analysis when run on 32 NVIDIA Grace Blackwell superchips inside a GB200 NVL72 system. The release lands alongside DAQIRI, a streaming library for live detector data, and ALCHEMI, a chemistry and materials microservice suite.
The numbers matter because the instruments behind them generate data faster than legacy CPU pipelines can absorb. The LSST camera, the largest digital camera ever built, captures billions of distant galaxies and faint nearby objects, and traditional pipelines were a bottleneck between raw capture and usable science. cuPhoton, developed with Princeton University researchers and slated for use at Harvard University as well, is designed to load, process, analyze, and visualize petabyte-scale multidimensional datasets from telescopes, X-rays, and laser experiments.
DAQIRI — short for Data Acquisition for Integrated Real-time Instruments — addresses a different choke point. Older acquisition systems are bound to fixed hardware and drop data when sensors out-produce storage. DAQIRI handles the stream as it arrives and pipes it into NVIDIA software for inference at line rate. The first showcase is A-GHOST, a project from CERN, the University of Chicago, and University College London built under CERN openlab, which uses DAQIRI to run AI in real time on ATLAS collision data.
Key facts
- 01cuPhoton accelerated loading and reading of Rubin Observatory FITS images by 14,900x in early access.
- 02Signal processing ran up to 8,400x faster on 32 NVIDIA Grace Blackwell superchips.
- 03ALCHEMI cut high-throughput materials screening 50x and trimmed magnetic properties calculation by 30% for Lila Sciences.
- 04A-GHOST uses DAQIRI to analyze ATLAS collision data that would otherwise be discarded — over 99% of the stream.
- 05ALCHEMI's TensorNet kernels delivered a 6x training and inference speedup with 3x lower memory use.
That use case has a concrete payoff. ATLAS rejects more than 99% of its collision data due to storage constraints, meaning rare signals get thrown away before any human looks at them. A-GHOST analyzes that rejected stream, catching potentially interesting events that would otherwise be lost. It is a representative pattern across modern experimental science: the cost is no longer collecting the data, it is keeping up with it.
ALCHEMI is the materials and chemistry counterpart. NVIDIA released two ALCHEMI NIM microservices in March at NVIDIA GTC San Jose — one for batched geometry relaxation, one for batched molecular dynamics — that let researchers simulate millions of candidate molecules at once. A third microservice wrapping the Vienna Ab initio Simulation Package, expected later this summer, runs multiple VASP calculations on a single GPU through NVIDIA Multi-Process Service, posting a 3x speedup on geometry optimization.
Lila Sciences, which is building an autonomous lab platform for life sciences, chemistry, and materials, ran the most complete public test of the stack to date. Using the ALCHEMI NIM microservice for batched geometry relaxation, the company accelerated high-throughput materials screening by 50x, then layered on the ALCHEMI VASP microservice in early access to compute magnetic properties for shortlisted candidates 30% faster. ALCHEMI's specialized kernels for TensorNet contributed a further 6x speedup in training and inference and cut memory use 3x, collapsing what had been weeks of simulation into days.
Lila Sciences is also pulling in adjacent NVIDIA pieces — Megatron-LM and Nemotron for training, including the Nemotron 3 Nano and Nemotron 3 Super open models, plus BioNeMo for molecular generation, Triton and NIM for inference serving, and Omniverse for digital twins. The architecture is a vertical stack from foundation models to physics simulation to lab automation, with each layer feeding the next.
“The work showcases using a powerful computing stack assembled to accelerate discovery at a scale no individual scientist could achieve alone.”— Andy Beam, Cofounder and CTO of Lila Sciences
The launches fit a pattern NVIDIA has been building across scientific computing. Earlier this month the company highlighted JUPITER, Europe's first exascale supercomputer, posting results across brain, climate, and quantum workloads, and the NAIRR pilot crossed 700 research projects running on DGX infrastructure. Domain-specific libraries — cuPhoton for astronomy, DAQIRI for detector streams, ALCHEMI for materials — extend that footprint from raw compute into ready-to-run scientific pipelines.
The caveat is access and integration. cuPhoton is still expected this summer, the ALCHEMI VASP microservice is in early access, and the most dramatic speedups assume GB200 NVL72-class hardware that most academic groups do not own outright. Speedup numbers from a vendor benchmark also reflect well-tuned reference workloads; real-world deployment at observatories and national labs will reveal how much of the 14,900x carries over to messier production data. DAQIRI is on GitHub now, and the ALCHEMI Toolkit ships through PyPI and the NVIDIA NGC catalog.
The strategic point is that NVIDIA is no longer selling only GPUs to science — it is selling the data path. cuPhoton owns the telescope-to-insight pipeline, DAQIRI owns the detector-to-model pipeline, ALCHEMI owns the simulation-to-candidate pipeline, and Nemotron and BioNeMo sit on top as the model layer. For research institutions deciding what to build experiments around for the next decade, the cost of switching off this stack rises with every new library. That is the durable moat, and it is being laid one scientific domain at a time.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
