In the mean time, Tesla depends on a supercomputer consisting of Nvidia A100 GPUs, which consists of 5,760 GPU models organized throughout 820 nodes, able to churning out 1.8 Exa-FLOPS. Nonetheless, the Dojo can leapfrog these figures simply. Tesla has beforehand claimed that only some Dojo cupboards can carry out the identical type of computerized labeling work as 1000’s of GPUs clustered collectively. However as a substitute of outsourced cores, Dojo depends on a self-designed D1 chip that’s manufactured by TSMC based mostly on its 7nm fabrication course of. Fairly than using centralized server-center-inspired structure, Dojo shall be a scalable machine that may finally obtain the shape referred to as ExaPODs.
Every D1 chip accommodates over 300 computing cores, whereas the D1 chips themselves are clustered to kind tiles. A group of six tiles would represent a single System Tray, whereas a pair of System Trays would occupy one cupboard. Tesla’s engineers would then mix 10 such cupboards to kind one ExaPOD. Tesla provided an in depth have a look at Dojo at its AI Day occasion final yr.
Dojo goals to considerably speed up the auto-labeling course of that goes into coaching the Autopilot mannequin, permitting it to extra precisely acknowledge real-world objects and comprehend conditions in order that it may accordingly make the precise choice. However even in the course of the early take a look at section, Dojo proved to be so highly effective that it reportedly overloaded the native energy grid.