Unverified50% confidenceTradeoffExact time
PD separation breaks data locality, as the KV cache generated by prefill nodes needs to be transferred to decode nodes over the network, with bandwidth consumption and latency increasing as input sequence length grows
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified寄存器溢出（Register Spilling）会将溢出变量存入本地显存，引入额外全局显存访问导致性能倒退68% similarUnverifiedOne of the core bottlenecks in long-context inference is memory bandwidth, because the KV cache grows linearly with context length and must be read from memory in real time.67% similarUnverified数据在服务器端进行推理时必然处于明文状态，GPU显存中的激活值、KV缓存以及批处理队列中的请求数据都可能成为侧信道攻击目标61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945825API
curl https://kongchang.com/api/v1/knowledge/claims/945825MCP
get_claim(id=945825)