Unverified70% confidenceFactTime unknown
数据中心内部的GPU互联带宽通常在100-400 Gbps量级（通过InfiniBand HDR/NDR或NVSwitch），而数据中心之间的广域网带宽可能仅有1-10 Gbps
2
Sources
70%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Decoupled DiLoCo详解：让大规模GPU训练不再怕宕机
twitterJeffDean
Related Entities
Related Claims
Verified在H100 NVLink架构中，单台服务器内8块GPU可通过NVSwitch全互联，总带宽达900GB/s，相比PCIe 5.0的64GB/s有数量级提升73% similarVerified传统x86平台上PCIe总线连接独立GPU与主板内存的带宽上限约为64GB/s72% similarVerified3080Ti的显存带宽为912 GB/s，远高于CPU与GPU之间PCIe总线带宽的16-32 GB/s72% similarUnverified8卡DGX B200系统的GPU间总带宽达到14.4TB/s72% similarUnverifiedGPU HBM带宽通常在2-3 TB/s量级，H100为3.35 TB/s，而PCIe 5.0连接的Host DRAM有效带宽仅约64 GB/s，两者相差近50倍71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/22916API
curl https://kongchang.com/api/v1/knowledge/claims/22916MCP
get_claim(id=22916)