Verified65% confidenceTradeoffExact time
分层加载导致推理速度受限，其根本原因是 PCIe 总线带宽（PCIe 4.0 理论带宽约 32GB/s）远低于 GPU 内部显存带宽
3
Sources
65%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
Related Claims
Unverified大模型推理速度往往不受限于算力（FLOPS），而受限于内存带宽；PCIe 4.0 x16理论带宽约32GB/s，GPU显存带宽（GDDR6X约1TB/s）相差30倍以上84% similarUnverifiedPCIe 4.0 x16带宽约64GB/s，且每次传输需要显式的cudaMemcpy调用，在推理延迟敏感场景中是显著瓶颈83% similarVerifiedPCIe 4.0 x16的理论带宽约为32GB/s81% similarUnverifiedPCIe万兆网卡多需PCIe x4插槽带宽，插在x1槽会限速；选购前确认主板有空闲x4及以上物理槽位与足够lane81% similarUnverifiedPCIe万兆网卡需要至少PCIe 2.0 x4或3.0 x2带宽，插在x1插槽会被限速无法跑满万兆，安装前确认主板插槽物理与带宽规格79% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/563112API
curl https://kongchang.com/api/v1/knowledge/claims/563112MCP
get_claim(id=563112)