Artificial intelligence news.
Straight from the original announcements.

NVIDIA Says cuPhoton Correlated a 10 TB EHT Dataset in 70 Seconds

November 13, 2026, 22:17By the AI News Bites Editorial TeamSource: NVIDIA Developer
NVIDIA Says cuPhoton Correlated a 10 TB EHT Dataset in 70 Seconds
Image: NVIDIA Developer

NVIDIA says its cuPhoton.xBeam software processed a 10 TB radio astronomy dataset in 70 seconds on a system with 64 GPUs. The result, according to NVIDIA, was about 460 times faster than an x86 CPU baseline, highlighting how GPU correlation could accelerate data processing for black hole imaging.

Key points
  • cuPhoton.xBeam is NVIDIA’s GPU-based FX correlator for radio interferometry.
  • NVIDIA says it processed a 10 TB dataset in 70 seconds.
  • The test used an NVIDIA Vera Rubin NVL72 system with 64 GPUs.
  • NVIDIA reports about a 460x speedup over a 31,992-second x86 CPU baseline.
  • xBeam outputs visibilities for downstream calibration and image reconstruction.

GPU processing for the Event Horizon Telescope

cuPhoton.xBeam is designed to perform radio interferometry correlation for the Event Horizon Telescope (EHT). The EHT links 13 radio observatory stations across locations including Hawaii, Chile, Spain, Mexico and the South Pole. These stations record millimeter and submillimeter radio signals as high-speed voltage streams, which are later synchronized and combined.

NVIDIA says the signals are converted into visibilities, measurements that provide inputs for calibration and image reconstruction. The resulting data can help researchers reconstruct images of objects that are too distant and small for a single telescope to resolve, including black holes surrounded by hot, glowing gas.

What cuPhoton.xBeam does

xBeam is the FX correlator in the cuPhoton toolkit. It first restructures recordings for parallel reads and builds a .vdix index that records the location of each VDIF frame. This allows later reads to jump directly to required data without rewriting the recording.

The software then applies delay and fringe corrections so signals from different stations line up, channelizes voltage streams with a fast Fourier transform, and cross-multiplies station and polarization pairs. It accumulates those products over the integration time, tracks invalid input and writes the resulting visibilities in either DiFX’s SWIN format or xBeam’s xbeam-v1 format.

According to NVIDIA, these operations run on GPUs. The company says xBeam can split observation time across GPUs on one node with --devices, or across multiple nodes with --shard-count and --shard-index. The merge-shards function combines partial SWIN outputs into a complete visibility set.

A 460-fold speedup in NVIDIA’s test

NVIDIA reports that xBeam processed a 10 TB dataset in 70 seconds on an NVIDIA Vera Rubin NVL72 system with 64 GPUs and GPUDirect Storage in compatibility mode. The company compares that result with 31,992 seconds, or 8.89 hours, for an x86 CPU baseline, describing the GPU result as about 460 times faster.

The processing stage is an early part of the imaging workflow. NVIDIA says the resulting visibilities serve as inputs to downstream calibration and image reconstruction, which are outside the scope of the tutorial. The cuPhoton toolkit also does not currently include a separate component for constructing a 3D gas model from those measurements.

NVIDIA is collaborating with the Harvard-Smithsonian Center for Astrophysics and the MIT Haystack Observatory on data ingestion and correlation. The company presents the work as relevant to the next-generation EHT, which aims to add observing sites and improve image quality while tracking changes around black holes over time. NVIDIA also points to potential studies of rapidly changing sources such as Sgr A*, where evolving signals make reconstruction more difficult.

Related news

Dell expands AI Factory with storage, compute and governance tools
Chips and infrastructureNovember 16, 2026, 07:00

Dell expands AI Factory with storage, compute and governance tools

Dell Technologies is expanding its Dell AI Factory with NVIDIA with persistent storage, additional compute options and supported software aimed at moving agentic AI from…

NVIDIA cuBLAS adds FP64 algorithms for faster scientific simulations
Chips and infrastructureNovember 13, 2026, 23:13

NVIDIA cuBLAS adds FP64 algorithms for faster scientific simulations

NVIDIA has added two algorithms to cuBLAS that accelerate double-precision matrix multiplication on integer Tensor Cores while preserving full FP64 accuracy. The company…

Hagens Berman Updates NVIDIA Investor Lawsuit After Reported CEO Resignation
BusinessNovember 13, 2026, 18:49

Hagens Berman Updates NVIDIA Investor Lawsuit After Reported CEO Resignation

Hagens Berman says it has updated its investigation into a pending securities class action involving NVIDIA after the reported resignation of CEO Jensen Huang. The firm…