Tesla Dojo Supercomputer Tile — Smart and Beautiful

This one tile can do 9 quadrillion floating point calculations per second, optimized for AI training. I saw it up close today.

Each 15kW tile holds a 5×5 array of blocks, each with 12 D1 multi-chip-module stacks, with RF shielding between them. These are custom AI chips by Tesla, with 362 teraFLOPS per block. Connectors around the perimeter provide 36Tb/s of inter-tile bandwidth.

The Dojo ExaPod supercomputer has 120 of these tiles, for 1.1 exaFLOPS of AI-optimized compute.