On November 16, 2026, NVIDIA announced an expansion of its CUDA-X library family, including cuPhoton, cuPotential, and cuSpa. In its announcement blog post, the company claims it cut black hole observation data analysis time from "more than a month to 8 minutes." However, that figure does not specify the data size or hardware configuration, and a separately published NVIDIA technical blog post presents measured results for a more narrowly defined workload. Evaluating the announcement requires distinguishing which steps were accelerated and under what conditions, and checking how each library, from drug discovery to spatial omics, has been validated.

AD

Radio Interferometer Correlation and the Role of xBeam

An NVIDIA technical blog post dated November 13, 2026 includes concrete benchmark results for xBeam, a module within cuPhoton. On a Vera Rubin NVL72 system with 64 GPUs and GPUDirect Storage enabled in compatibility mode, the full pipeline of reading, preparing, and correlating a 10 TB dataset reportedly finished in 70 seconds. The same job on an x86 CPU took 31,992 seconds (8.89 hours), a speedup of roughly 460x.

According to NVIDIA, the cuPhoton team is working with the Harvard-Smithsonian Center for Astrophysics and MIT Haystack Observatory to accelerate data ingestion and correlation with xBeam. However, no peer-reviewed paper or independent third-party reproduction of the cuPhoton or xBeam benchmarks has been identified at this time. The published performance figures are engineering measurements by NVIDIA.

Observing the outline of a black hole requires a virtual telescope the size of the Earth. The Event Horizon Telescope (EHT) links 13 radio observatories in locations including Hawaii, Chile, Spain, Mexico, and Antarctica into a very-long-baseline interferometry (VLBI) array. Each telescope records two polarizations at up to 64 Gbps, and the observation data reaches the petabyte scale.

The "correlation" that xBeam accelerates is the computation of complex visibilities for each combination of telescope pair, frequency channel, and short time interval. Visibilities are Fourier components of the sky's brightness distribution, sampled at a limited number of observing points. Reconstructing a black hole image from this sparse data requires solving an inverse problem, separate from correlation.

xBeam, part of cuPhoton v0.1.3, which is open source on GitHub, runs FX-style correlation on the GPU. It performs delay and fringe corrections, splits the signal into frequency channels via fast Fourier transform (FFT), then cross-multiplies and accumulates, writing the results in DiFX's SWIN format or the proprietary xbeam-v1 format.

In VLBI, the standard tool at observatories worldwide has been DiFX, a CPU-based open-source FX software correlator developed from 2005 and released publicly in 2007 (Deller et al. 2011, PASP 123 275, DOI: 10.1086/658907). NVIDIA's technical blog also uses DiFX as the CPU reference correlator to validate xBeam's results. Downstream calibration and imaging after correlation are handled by specialized tools such as HOPS, rPICARD, and eht-imaging.

In other words, xBeam is responsible for applying the delay values specified in the job file and generating visibilities. It does not calculate delays from station positions or trace the paths of light around a black hole. The 3D gas model reconstruction via Gaussian splatting described in the technical blog is also an independent downstream process and is not built into the current cuPhoton. The result NVIDIA mentioned in its announcement blog, successfully correlating observation data of the black hole jets J1924 and 3C279, likewise concerns this correlation step.

Comparing earlier correlator research with xBeam reveals differences not only in hardware but also in what was evaluated and on what basis.

Approach Hardware Reported performance and scope Nature and date of evidence
DiFX-2 (CPU correlator) General-purpose CPU cluster FX correlation (visibility generation) with a track record in VLBI observations worldwide Peer-reviewed paper (PASP, 2011)
Tensor-Core GPU X-engine (earlier prototype) 2 servers, 4 GPUs Microbenchmark stage; a full correlation and beamforming system not yet built Peer-reviewed design study (Galaxies, 2023)
cuPhoton.xBeam (NVIDIA GPU correlator) Vera Rubin NVL72 (64 GPUs) 70 seconds from reading 10 TB of data through correlation; about 460x faster than 31,992 seconds on x86 CPU NVIDIA technical blog (November 2026)

DiFX has a long operational track record, while the Tensor-Core GPU X-engine is an academic-stage prototype. xBeam's roughly 460x figure, meanwhile, is an NVIDIA measurement using a specific GPU system and dataset, so these cannot be treated as a like-for-like performance comparison.

Surging Observation Data and How to Assess the Speedup

The EHT's need for faster correlation stems from the rapid growth in observation data. According to a paper on the 2017 observations of the blazar J1924-2914 (Issaoun et al. 2022, ApJ 934:145, DOI: 10.3847/1538-4357/ac7a40), the conventional workflow involved physically shipping the data recorded at each station on storage media to MIT Haystack Observatory or the Max Planck Institute for Radio Astronomy in Bonn, Germany, for correlation.

A design study for the next-generation expansion, ngEHT (Galaxies 2023, 11(1):13, DOI: 10.3390/galaxies11010013, published January 11, 2023), projects that as sideband bandwidth grows from 4 GHz to 8 GHz, about 21 stations will each record data at 256 Gbps. Annual observation days are expected to rise from 7 to 60, and the data generated will reach tens of petabytes per year. This calls for dedicated, high-throughput correlators. The same paper built a prototype Tensor-Core GPU X-engine using two servers and four GPUs, but it stopped short of building a full system including correlation and beamforming.

xBeam scales performance by splitting scan data along the time axis and distributing the work across multiple GPUs. According to NVIDIA's technical blog, read bandwidth for the 10 TB dataset using GPUDirect Storage reached 20 GB/s with 4 GPUs and 170 GB/s with 64 GPUs. Within the measured range, performance improves as GPUs are added.

When assessing speedup figures, however, individual operations must be distinguished from whole-pipeline performance. A separate cuPhoton technical blog NVIDIA published on October 7, 2026 describes speedups in areas such as image loading and signal processing, but those figures apply to individual operations. The publicly available materials we could find do not give the specific before-and-after processing times, and they state that performance varies with workload, dataset, implementation, and hardware.

For xBeam's 70-second measurement as well, the comparison baseline is described only as "x86 CPU," and details such as the core count, software configuration, and code path behind the 31,992-second CPU run have not been disclosed. The roughly 460x figure should therefore be read as the performance gap under these specific measurement conditions, and there is no basis for assuming the same speedup on typical CPU environments.

AD

Performance and Accuracy of the cuPotential Molecular Dynamics Library

cuPotential, announced for drug discovery and materials design, is a CUDA-X library that accelerates the calculation of long-range non-bonded interactions in molecular dynamics (MD) simulations. In classical force field simulations with periodic boundary conditions, electrostatic and dispersion interactions are the main bottleneck. According to an NVIDIA technical blog dated November 16, 2026, this release also includes a plugin for OpenMM.

In molecular dynamics, particle mesh Ewald (PME) has been the de facto standard for about 30 years. Alternatives such as u-series, midtown splines, and prolates have been proposed, but their implementation and operational complexity have kept them from becoming widespread. NVIDIA worked with D. E. Shaw Research and the Flatiron Institute's Center for Computational Mathematics to develop the new -continuous splitting and diffusion functions that underlie cuPotential.

Benchmarks of isothermal-isobaric (NPT) ensembles in OpenMM measured performance on systems of 24,000 to 400,000 atoms. The test systems included DHFR, TYK2, BACE, ApoA1, 100,000 atoms of water, and cellulose. After a 500-step warmup, 3,000 steps were measured, and under conditions using cuPotential's own force calculation, integrator, and barostat, an end-to-end speedup of up to 3x was reported. However, the absolute processing times for the baseline and the comparison are not given in the publicly available materials we could find.

Beyond computational speed, NVIDIA also tested numerical accuracy. In a 10 ns rigid-water simulation with a 9 Å cutoff, a 2 fs time step, and a temperature of 300 K, compared against a double-precision reference, cuPotential reportedly showed less energy drift than PME with default settings. However, the specific numbers for this comparison are not given in the publicly available materials we could find.

The current OpenMM plugin also has functional limitations: it does not support LJPME, triclinic simulation boxes, or OpenMM's particle parameter and exception parameter offsets. These evaluations address simulation throughput and numerical error, not experimental validation of drug candidates or clinical efficacy. Adoption by multiple pharmaceutical-related companies has also been announced, but what has been shown so far is statements about adopting the tool.

Expansion into Spatial Omics and Fusion, and the Role of AI Agents

The CUDA-X expansion also extends to single-cell biology and fusion research. cuSpa, for spatial omics analysis, is open source under the Apache-2.0 license and uses a GPU kernel for point-in-polygon tests to assign transcript coordinates to cell polygons. This builds a sparse cell-by-gene count matrix.

According to NVIDIA's announcement blog, at Memorial Sloan Kettering Cancer Center, cuSpa achieved a 900x speedup over CPU for generating spatial graphs for patient and cancer models. However, the specific before-and-after processing times are not given in the publicly available materials we could find. In addition, rapids-singlecell, the single-cell analysis framework developed jointly with scverse, is now offered as a single container that supports AI agents.

In fusion, a case was presented in which researchers at Pacific Fusion use cuPhoton.xLens to extract information in real time while a plasma simulation is running. The configuration shown can also work with an NVIDIA Nemotron-based AI agent running locally on the same system when needed.

NVIDIA positions the announcement as part of an effort to support accelerated scientific discovery through AI agents. However, the benchmarks reported in the xBeam technical blog target GPU compute acceleration and do not measure autonomous control or orchestration of an entire pipeline by AI agents. Such capabilities are at the stage of proposed future applications.

The source code for cuPhoton, cuPotential, and cuSpa is all published on GitHub, allowing third parties to verify the implementations and performance. What will be assessed next is whether the speedups shown can be reproduced in other research environments and datasets, and how much benefit integration with AI agents actually brings to real research workflows.