NVIDIA brings new AI-for-science software to ISC, aiming to make data-intensive research run in “real time”

NVIDIA used the International Supercomputing Conference in Hamburg to roll out a new set of software aimed at speeding up some of the most computation-heavy parts of modern science, from telescope data processing to chemistry and materials modeling. The message was that as research instruments generate more data, the bottleneck is increasingly not measurement but the speed at which computers can ingest, analyze, and iterate on that data. At the center of the announcement is the NVIDIA DAQIRI library, short for Data Acquisition for Integrated Real-time Instruments. NVIDIA described DAQIRI as high-performance networking software designed to stream data from fast detectors and sensors into GPU-accelerated pipelines without dropping readings. The company said older approaches tied to fixed hardware can lose data when instruments produce it faster than storage can keep up, while DAQIRI is built to handle the stream as it arrives. NVIDIA also introduced work that it says extends this “accelerate the pipeline” idea into several scientific domains through its CUDA-X software stack, a collection of tools and libraries that target higher performance across application areas. Among the new items are NVIDIA ALCHEMI NIM microservices, which NVIDIA described as domain-specific building blocks for accelerating chemical and materials discovery, and a reference code called NVIDIA cuPhoton intended for astrophysics workflows. NVIDIA said the components are meant to turn time-consuming CPU workloads into GPU-accelerated processes that can run with near real-time responsiveness. In astronomy, NVIDIA said cuPhoton is designed to load, process, analyze, and visualize multidimensional data from telescopes and other instruments, including petabytes of observations stored in standard astronomical file formats. The company pointed to FITS, the Flexible Image Transport System, as a common file format used by observatories. NVIDIA said cuPhoton is built as a reference code for scientists extracting insights from that data, and it is intended to support end-to-end accelerated pipelines when paired with other CUDA-X technologies. The company’s performance examples focused on the Rubin Observatory’s Legacy Survey of Space and Time, or LSST, which is expected to produce extremely large volumes of astronomical imagery over time. NVIDIA said early access results on NVIDIA GB200 NVL72 systems accelerated FITS loading and reading of Rubin LSST images by 14,900 times. It also said cuPhoton enabled up to 8,400 times faster announcement processing and analysis using 32 NVIDIA Grace Blackwell superchips. NVIDIA framed those speedups as a practical acceleration for interpreting what the LSST camera sees, including images of billions of distant galaxies and fainter, nearby objects that produce comparatively weak indicates. NVIDIA said the underlying goal is faster insights from the LSST camera and similar observatory pipelines, where scientific conclusions depend on how quickly data can be processed, filtered, and studied. Outside astronomy, NVIDIA said DAQIRI has been used in a particle physics research project called A-GHOST, developed by researchers from CERN, the University of Chicago, and University College London under CERN openlab. NVIDIA said the system uses DAQIRI to run AI in real time on collision data recorded by the ATLAS experiment at CERN, analyzing event data that would normally be rejected. NVIDIA said more than 99% of the data is typically rejected because of storage constraints, and that A-GHOST aims to catch potentially interesting indicates that might otherwise be lost. For materials and chemistry, NVIDIA described its ALCHEMI offering as a set of microservices and a toolkit intended to accelerate the discovery loop, including tasks that simulate how molecules and materials behave. NVIDIA said it released in March two ALCHEMI NIM microservices for batched geometry relaxation (BGR) and batched molecular dynamics (BMD). It explained BGR as a way to find stable molecular or material structures and BMD as simulating how those entities move over time. The company said the approach is built to simulate millions of molecules and materials at once. NVIDIA also said ALCHEMI is expected to soon include a microservice for VASP, the Vienna Ab initio Simulation Package, which is widely used for materials simulations. NVIDIA said the VASP microservice is designed to increase GPU throughput by running multiple VASP calculations on a single GPU with NVIDIA Multi-Process Service, a feature for time-sharing GPU resources across workloads. The company stated that in testing it achieved a 3x speedup for geometry optimization, the step of finding the most stable arrangement of atoms. The announcement included collaboration examples, including Princeton University and Harvard University for cuPhoton. NVIDIA said Princeton helped develop cuPhoton and will use it, and that both Princeton and Harvard plan to use it for processing and analysis of massive data collected by observatories and dark energy surveys. NVIDIA also highlighted Lila Sciences, which the company described as building a scientific superintelligence platform and autonomous lab for life sciences, chemistry, and materials science. NVIDIA said Lila used ALCHEMI to accelerate high-throughput materials screening by 50x using the BGR microservice, then sped up magnetic property calculations by 30% for shortlisted candidates using an early-access VASP microservice. NVIDIA added that specialized kernels reduced memory usage by 3x and that speedups compounded, reducing simulations previously taking weeks to just days. Availability and what’s next were also part of the update. NVIDIA said ALCHEMI NIM microservices can be downloaded from the NVIDIA NGC catalog, and that the ALCHEMI Toolkit and Toolkit-Ops are available for download from GitHub and PyPI. The company said the ALCHEMI NIM microservice for VASP is expected later this summer, and it said cuPhoton is expected later this summer as well. NVIDIA also said DAQIRI is now available on GitHub. Even with the performance figures, there are limits to what can be concluded from this kind of announcement. NVIDIA did not specify, in the published briefing, the full hardware and dataset configurations behind each headline multiplier, how broadly transferable the results are to other observatories or instruments, or whether results are consistent across all pipeline stages beyond the workloads it highlighted. For example, the LSST FITS acceleration was described for early access, and the company did not provide independent benchmarking methodology or third-party validation details in the same release. Those qualifiers matter because end-to-end science pipelines often include additional steps, including calibration, quality control, and data management, not only file ingest and core computation. What to watch next is whether NVIDIA’s “accelerated pipeline” approach shows up in broader deployments beyond early-access environments, including whether additional astronomy workflows adopt cuPhoton as an industry reference. On the software side, scientists and developers will likely pay close attention to the expected timing of the ALCHEMI VASP microservice and to the rollout of DAQIRI and cuPhoton for teams building real-time and high-throughput analysis systems. If the performance gains hold across larger operational settings, the bigger implication is that access to accelerated computing could translate into faster iteration cycles in discovery, not just faster computation.

keyFacts

NVIDIA unveiled new AI-for-science software at ISC in Hamburg, including the DAQIRI data-streaming library, ALCHEMI NIM microservices for chemistry and materials, and the cuPhoton reference code for astronomy workflows.

DAQIRI is designed to stream data from fast detectors and sensors into GPU software to reduce data loss when instrument output outpaces storage capacity.

NVIDIA said cuPhoton accelerates FITS (Flexible Image Transport System) image loading and reading for Rubin Observatory LSST data by 14,900x on NVIDIA GB200 NVL72 systems, and improved announcement processing and analysis by up to 8,400x using 32 NVIDIA Grace Blackwell superchips.

NVIDIA cited A-GHOST, an AI system for CERN’s ATLAS experiment, saying it uses DAQIRI for real-time analysis of collision data that would otherwise be rejected due to storage constraints.

In materials discovery, NVIDIA said ALCHEMI NIM microservices released in March include batched geometry relaxation (BGR) and batched molecular dynamics (BMD), and that a VASP microservice is expected later this summer.

The company said ALCHEMI Toolkit and Toolkit-Ops are available via GitHub and PyPI, ALCHEMI NIM microservices are available via NVIDIA NGC, DAQIRI is now on GitHub, and cuPhoton is expected later this summer.

whyItMatters

Faster AI-enabled pipelines can shift the bottleneck in data-intensive research from measurement to computation, potentially compressing discovery timelines in astronomy, particle physics, and materials science.

If early performance results translate to production deployments, NVIDIA’s CUDA-X-aligned approach could strengthen the case for GPUs as core infrastructure for scientific data processing, not only model training.

The announcement emphasizes real-time streaming and throughput, which could matter as next-generation instruments and surveys increase data rates faster than traditional compute and storage workflows can handle.

Specialized microservices such as ALCHEMI for VASP could lower integration friction for researchers already using standard simulation tools while still tapping GPU acceleration.

sourceTrail

sourceTrail

sourceTrail must include the original source and researched URLs. Here are the URLs used:

sourceTrailEntries