Researchers Announce Self-Hosted GPU Shed for AI Workloads

Joseph Suarez posted that his team made their infrastructure 1000x faster so it runs on a potato. The change follows their purchase of GPUs that they now self-host in a shed. Andrew Carr answered a question about the number of GPUs by stating the setup ranks in the top 25 for compute capacity in Utah. Gowthami Somepalli asked about the GPU count and described a similar home shed approach that she said cut rental costs by ten times and enabled more experiments than cloud limits allow. The posts are replies among visible conversation on the platform.