Unverified50% confidenceFactExact time
PCIe Gen4 x16 bandwidth is approximately 64GB/s, which often becomes a hard bottleneck in memory offloading scenarios.
1
Sources
50%
Confidence
Long-term
Relevance
10/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCPU 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.80% similarUnverified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/s79% 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/s77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/974696API
curl https://kongchang.com/api/v1/knowledge/claims/974696MCP
get_claim(id=974696)