NVIDIA says GPUs correlated 10 TB of black hole data in 70 seconds

The latest:

A 10 TB radio-telescope dataset was correlated in 70 seconds on an NVIDIA Vera Rubin NVL72 system with 64 GPUs, according to NVIDIA, which put the x86 CPU reference run at 31,992 seconds, or 8.89 hours. The company said the work on its cuPhoton.xBeam correlator is being carried out with the Harvard-Smithsonian Center for Astrophysics and the MIT Haystack Observatory.

Details:

• The benchmark: NVIDIA said the 70-second run used GPUDirect Storage in compatibility mode across 64 GPUs, against a CPU baseline of 31,992 seconds. The post does not name the telescope array that produced the 10 TB dataset, or the specification of the x86 reference machine.
• What the tool does: According to NVIDIA, cuPhoton.xBeam is an FX correlator that multiplies signals from pairs of telescopes to produce one complex visibility per station pair, per frequency channel, per short time interval. Those visibilities are the measurements image reconstruction algorithms later work from.
• The five stages: NVIDIA describes the pipeline as ingest, delay and fringe correction, channelization, cross-multiplication and accumulation, then writing visibilities. Phase correction, the Fourier transform and the cross-multiply run on the GPU, with output written in DiFX’s SWIN format or xBeam’s own xbeam-v1 format.
• The data problem: The Event Horizon Telescope records radio signals in two polarizations at up to 64 gigabits per second per telescope, generating petabytes of data, according to the post. The array links 13 stations in locations including Hawaii, Chile, Spain, Mexico and the South Pole.
• Why imaging is hard: NVIDIA notes the ring observed around M87* spans just 42 microarcseconds, and that no physical lens forms the picture. Earth’s rotation traces only sparse elliptical measurement tracks, so reconstruction algorithms must solve an inverse problem from incomplete samples. Rapid changes in sources such as Sgr A* break methods that assume a static image.
• Stated limits: NVIDIA says xBeam does not compute delays from station positions but uses values supplied in the job file, and does not trace light paths around a black hole. The toolkit also does not currently include the Gaussian splatting step used to build 3D gas models from visibilities.
• Validation reference: The CPU correlator DiFX is used as the reference for validating xBeam results, according to the post, and downstream tools including HOPS, rPICARD and eht-imaging take the visibilities onward for calibration and imaging.
• The next arrays: The next-generation EHT aims to add observing sites to sharpen images and track changes around black holes over time. The proposed Black Hole Explorer is designed to measure the thin photon ring, and ngEHT observations at 345 GHz could detect its signature through polarization measurements.

Background:

Black holes emit no light themselves. Telescopes capture radio emission from superheated gas outside the event horizon, with the hole’s gravity bending that light into the bright ring framing the central shadow. Very-long-baseline interferometry synchronizes separated stations to reconstruct the image.

Between the lines:

The bottleneck NVIDIA is targeting is throughput, not optics. With stations recording up to 64 gigabits per second each and the array producing petabytes, correlation time sets how quickly raw voltages become usable visibilities. That matters most for Sgr A*, which changes during an observation, and for the ngEHT plan to add sites and track motion over time — both of which increase data volume before they improve the picture.

What’s next

Watch whether cuPhoton benchmarks hold on live EHT observing campaigns rather than a single 10 TB dataset, whether ngEHT adds stations at 345 GHz, and whether the proposed Black Hole Explorer mission advances toward measuring the photon ring.