Everyone knows Cerebras by their wafer scale chips. The less understood part is the 12TB of external memory. That's the real reason why large models fit by defa

Everyone knows Cerebras by their wafer scale chips. The less understood part is the 12TB of external memory. That's the real reason why large models fit by default and you don't have to chop it up in software ala megatron/deepspeed.