InfiniBand
科技
网络
The low-latency training fabric NVIDIA acquired with Mellanox.
您正在查看一条技术记录：AI 硬件的一个基础构件——它的功能、谁来制造，以及哪些芯片和系统使用它。
InfiniBand's RDMA transport, in-network reduction (SHARP) and lossless design made it the default scale-out network for big training clusters — Microsoft's OpenAI systems and thousands of others run Quantum-2 400 Gb/s fabrics, with 800 Gb/s Quantum-X families following. Its share is now contested by AI-tuned Ethernet, which costs less and has a broader vendor base.
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。
维护方
来源
记录审阅截至 2026-01-15 — 查看 方法论.