[cgdl]

I'd say llm inference requires both memory capacity and bandwidth. Cerebras provides bandwidth with on-chip SRAM, but not capacity (an entire wafer has only 44GB SRAM).