from ml-engineering-master\network\README.md
"Other intra-node solutions typically have the same all-to-all and peer-to-peer intra-node bandwidth, so Infinity Fabric appears to be dramatically slower. I suppose that is because these were created mainly for inference, as these slow speeds would dramatically slow down LLM training."
Why do you suggest that AMD’s lower point-to-point bandwidth makes its intra-node interconnect more suitable for inference rather than training? What is the basis for this conclusion? Since both training and multi-GPU inference can involve frequent GPU-to-GPU communication, especially with tensor parallelism, I would like to understand why the lower P2P bandwidth would affect training more significantly or lead to this distinction. Is this conclusion supported by specific benchmarks, workload characteristics, or a particular parallelism strategy?
from ml-engineering-master\network\README.md
"Other intra-node solutions typically have the same all-to-all and peer-to-peer intra-node bandwidth, so Infinity Fabric appears to be dramatically slower. I suppose that is because these were created mainly for inference, as these slow speeds would dramatically slow down LLM training."
Why do you suggest that AMD’s lower point-to-point bandwidth makes its intra-node interconnect more suitable for inference rather than training? What is the basis for this conclusion? Since both training and multi-GPU inference can involve frequent GPU-to-GPU communication, especially with tensor parallelism, I would like to understand why the lower P2P bandwidth would affect training more significantly or lead to this distinction. Is this conclusion supported by specific benchmarks, workload characteristics, or a particular parallelism strategy?