Architecture · Rack
Chip to Rack
Start inside one Rubin die, a single piece of silicon, and pull back, step by step, until you are looking at a whole rack and the fabric that joins it to the rest of the cluster.
Die
A die is one piece of silicon cut from a wafer: the chip itself, before it is packaged. This Rubin GPU (graphics processing unit) die is packed with streaming multiprocessors, the GPU’s compute blocks, each holding tensor cores that do the matrix maths at the heart of every model.
Lower-precision number formats let each pass of the tensor cores do more useful work, which is why vendors now quote performance in very low precision.
During an agent turn: Every token the agent thinks is computed here.