SpaceX's self-developed C-language AI training framework is nearing completion and will be used to train Grok v5.

According to ME AI , based on Beating's monitoring, Elon Musk revealed that SpaceX's internal team is nearing completion of version 1.0 of its self-developed ultra-large-scale AI training framework written in C. In terms of hardware mapping and parallel mechanisms, the new training stack can precisely adapt to an ultra-large-scale computing cluster composed of 220,000 NVIDIA GB300 accelerator cards and 800G network cards. To squeeze out the underlying computing power, the framework is designed to be extremely close to a bare metal underlying layer and deeply employs pipeline parallelism technology. Musk revealed that when dealing with ultra-large-scale training tasks, the pure C language underlying architecture can potentially improve the running speed by more than an order of magnitude (more than 10 times) compared to Google's mainstream high-level AI framework JAX. The new self-developed training stack will run on SpaceX's Colossus supercomputing cluster, directly serving the full training and iteration of the next-generation large model Grok v5. (Source: ME)