Everybody's talking about GPUs. I think they're watching the wrong box.Meta trained Llama 3 on up to 16,000 GPUs, and th...

Everybody's talking about GPUs. I think they're watching the wrong box.Meta trained Llama 3 on up to 16,000 GPUs, and their own engineers wrote that a single slow straggler "can slow down thousands of other GPUs." Not a dead one. A slow one.If you've ever worked a bridge call where every dashboard was green, and users were still yelling, you already understand the hardest problem in AI.Why AI is secretly a networking problem:https://ryanthomashuff.com/why-ai-is-a-networking-problem/#Networking #AI #RDMA

Read Original

Related