3 ms·
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 44G
by cgdl 1y ago
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).