The Vera C. Rubin Observatory will collect roughly 20 terabytes every night — compared with Hubble's 1–2 gigabytes a day. That flood of imagery, joined by James Webb and NASA’s Nancy Grace Roman telescope, is pushing astronomers to rely on GPUs and AI to process and analyze the flow; projects such as UC Santa Cruz’s Morpheus show how those pipelines are increasing demand for high-end accelerators.

Data volumes jump by orders of magnitude The next wave of astronomy missions will produce data at a scale that would have been hard to imagine a decade ago. Key figures cited in project briefings and public releases show how the load is changing: - Nancy Grace Roman telescope: likely to deliver roughly 20,000 terabytes across its mission life after its launch in September 2026. - James Webb Space Telescope: downlinks about 57 gigabytes of imagery each day. - Vera C. Rubin Observatory: expected to collect roughly 20 terabytes every night once its wide survey begins. By comparison, the Hubble Space Telescope typically returns only 1 to 2 gigabytes per day. The jump affects not just storage but compute: higher pixel counts, faster cadences and the need to flag rare or transient events all favor more parallel processing. Why GPUs are now central to sky surveys GPUs excel at parallel math for graphics and matrix operations — the same workloads used in modern machine learning — which is why astronomers are moving from CPU-based analyses to GPU-accelerated pipelines. Brant Robertson, a UC Santa Cruz astrophysicist who has collaborated with Nvidia for many years, says the field has evolved from studying a handful of objects to applying GPU-powered models across massive data sets. “There’s been this evolution [from] looking at a few objects, to doing CPU-based analyses on large scales of the data set, to then doing GPU-accelerated versions of those same analyses,” Robertson said. The result: teams can process far more sky area, probe fainter objects and run more realistic simulations in less time. Morpheus and the move to transformer-based models Morpheus, a deep-learning pipeline developed at UC Santa Cruz by Robertson and Ryan Hausen, offers a concrete example. The original system used convolutional neural networks to scan telescope images and identify galaxies; early deployments on James Webb data revealed unexpectedly large numbers of a certain class of disc galaxies. The team is re-architecting Morpheus to use transformer-based architectures — the same model class behind recent advances in large language models — to analyze larger sky areas at once and process inputs faster. Researchers are also testing generative models trained on telescope data to denoise images, reconstruct missing pixels and produce candidate targets for follow-up. Each detection, denoising and ranking step becomes cheaper and faster with GPU acceleration. How astronomy demand intersects with commercial GPU supply The push toward GPUs in astronomy comes at a moment when commercial demand for high-end accelerators is already large. Cloud providers, AI startups, chip makers and research institutions all compete for the same classes of GPUs, meaning astronomy’s growing needs are one more factor tightening the global supply of high-performance accelerators.

Related Articles

The Roman telescope's planned September 2026 launch is a near-term milestone that will add substantial data to the queue, and projects like Morpheus show astronomy's shift to GPU-accelerated AI is already squeezing the global market for high-end accelerators.

This article was created with AI assistance.