inference-stack

SpaceX is finalizing a custom AI training stack written in C, utilizing pipeline parallelism and 220k GB300 GPUs to achieve over an order of magnitude speed improvement, with plans to develop an inference stack for reinforcement learning.