Hi everyone. I built Cluster Atlas because I kept wanting to know what hardware is actually inside the world's largest computers, and the information was spread across vendor PDFs, operator documentation, and TOP500 table rows. TOP500 ranks machines by speed, and it doesn't describe how they are built. Cluster Atlas covers 64 HPC and AI systems and shows the architecture of each one. You can start at the network fabric and open a group of nodes. From there you can look at the CPUs and accelerators inside a single node. You can also put two systems next to each other, and you can export any diagram as SVG or PNG. The catalog has a filter for finding hardware by requirement rather than by rank. You can filter by GPU model, memory per GPU, CPU architecture, interconnect, and whether the operator publishes a way to apply for research access. Every number links to the source it came from. Where an operator has not published a specification, I mark the value as unknown rather than guessing, because a wrong number is worse than a missing one. Some parts are still rough, and I would rather say so here than have you find them. Accelerator model names are not fully normalized, so filtering for one GPU can miss systems whose operator labeled the same chip differently. I would like to know which systems you want added, and which specifications you look at first when you size up a machine. If you find a wrong number, tell me. I will fix it and link the correction.

Cluster Atlas

Your map to Supercomputing Clusters.

Your map to Supercomputing Clusters.