📄 中文摘要
通过使用 libfabric、DMA-BUF 和 HCCL，构建类似 RDMA 的性能，解决了云主机网络接口卡（NIC）在分布式训练中的瓶颈问题。该技术实现了高效的数据传输，显著提升了 Gaudi 处理器在云环境中的性能，恢复了分布式训练的可扩展性。这一创新使得在云计算平台上进行大规模深度学习训练变得更加高效，推动了 AI 领域的进一步发展。
突破主机内存瓶颈：Peer Direct 如何提升 Gaudi 的云性能
📄 English Summary
Breaking the Host Memory Bottleneck: How Peer Direct Transformed Gaudi’s Cloud Performance
The study presents a solution to the memory bottleneck in cloud environments by engineering RDMA-like performance over cloud host NICs using libfabric, DMA-BUF, and HCCL. This approach enables efficient data transfer, significantly enhancing the performance of Gaudi processors in cloud settings and restoring scalability for distributed training. This innovation facilitates more efficient large-scale deep learning training on cloud computing platforms, advancing the development of AI technologies.