OpenAI推出新的Multipath Reliable Connection(MRC)网络协议,旨在减少大型AI集群中的拥塞和故障相关减速,支持超大规模扩展至数十万GPU,以应对日益增长的算力需求。
The company says its new Multipath Reliable Connection protocol reduces congestion and failure-related slowdowns in giant AI clusters as hyperscalers push toward hundreds of thousands of GPUs.