Unverified50% confidenceTradeoffExact time
分层加载导致推理速度成为代价，需要频繁在存储设备与显存之间搬运数据，PCIe 4.0 x16 理论双向带宽约为 64GB/s
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
Unverified主机卸载用带宽换显存空间，代价是数据传输延迟；激活重计算用算力换显存，二者可互补配合76% similarUnverified大型语言模型推理需将数GB至数十GB模型权重通过PCIe总线从主机内存搬运至显存，受限于PCIe Gen4/Gen5物理带宽上限约64GB/s76% similarUnverifiedCPU offload places some Transformer layers in system main memory to be computed by the CPU, with data transferred over the PCIe bus (PCIe 4.0 x16 theoretical bandwidth is approximately 32 GB/s), while GPU memory bandwidth is typically in the range of hundreds to over a thousand GB/s, making the bandwidth gap the fundamental cause of the performance bottleneck.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/567132API
curl https://kongchang.com/api/v1/knowledge/claims/567132MCP
get_claim(id=567132)