The Event Horizon Telescope links radio dishes around the world, and turning their huge voltage recordings into usable measurements takes hours on normal processors.

A GPU tool, cuPhoton.xBeam, that lines up and compares signals from each pair of telescopes and can split the work across many GPUs.

Serving GLM-5.2 on long inputs hits limits in compute, memory and links between GPUs at once, so tuning one part is not enough.

A set of changes on eight MI355X GPUs: splitting tokens across GPUs, a smaller 8-bit memory cache, less data sent between chips, and draft-and-check decoding.

A compressed version of GLM-5.3 needs about 464 GB of GPU memory, more than the 252 GB on one GB300 GPU.

A setup that links two DGX Stations over a fast network and splits the model's weights between them, served through one API with vLLM.

Medical imaging pipelines spend a lot of time just reading huge slide images from disk, which leaves the GPU waiting.

A stack of open-source tools on AMD GPUs that reads slides, runs ready-made medical models and supports clinician annotation, without code changes.