SpaceX Develops Custom AI Chip Rack, Bypassing NVIDIA Compute Boards to Boost Procurement Autonomy
4 hour ago / Read about 0 minute
Author:小编   

SpaceX is developing a custom version of the Vera Rubin NVL72 rack, called Kessel, which replaces the original NVLink cable cartridges with a passive PCB midplane while retaining 72 GPU packages, 18 compute trays, and 9 switch trays, with the trays arranged back-to-back around the midplane. This rack is a smaller-scale version of Kyber with a reduced scale-up domain, and is targeted for both ground and orbital deployment. Kessel uses an 84-layer M10 midplane, which is easier to manufacture and more feasible than the previously planned Kyber, which uses a 156-layer M9 midplane but requires higher-grade materials. Meanwhile, SpaceX has adjusted its component procurement model: it purchases chip packages directly from NVIDIA and sources the remaining components itself, bypassing NVIDIA's L6 compute boards to gain greater autonomy in supplier selection and component procurement.