Unverified50% confidenceBenchmarkExact time
Megatron-LM实验数据表明，NVLink互联下8卡张量并行可达7.5×以上线性加速，而PCIe 3.0 x16环境下7B模型通常仅达2.5-3×，超过4卡后通信开销完全抵消额外算力
1
Sources
50%
Confidence
Long-term
Relevance
7/22/2026
First Seen
Sources
Related Claims
Unverified多卡训练必须关注NVLink而非仅看单卡算力，PCIe 4.0互联带宽约64GB/s，NVLink可达600-900GB/s，8卡分布式训练时PCIe互联会导致AllReduce通信成为瓶颈，实际利用率可能不到50%72% similarUnverified接口速率是链路瓶颈：USB 3.2 Gen2约1000MB/s、雷电3/4约2800-3000MB/s、内置PCIe4.0 NVMe可达7000MB/s，外置SSD速度受接口上限制约而非仅取决于盘本身71% similarUnverified传输速率标称需看清：USB 3.2 Gen1=5Gbps、Gen2=10Gbps，很多商品用USB 3.0/3.1模糊表述实为5Gbps，接NVMe SSD想跑满速需认准10Gbps及以上71% similarUnverifiedGB300 搭载 NVLink 第五代互联技术，单机8卡的卡间带宽达到1.8TB/s70% similarUnverified全屋千兆NAS+多设备场景，交换机背板带宽应≥端口数×端口速率×2（全双工），如8口千兆需≥16Gbps无阻塞背板，标注'非阻塞/wire-speed'才能真正跑满70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/593791API
curl https://kongchang.com/api/v1/knowledge/claims/593791MCP
get_claim(id=593791)