Unverified50% confidenceBenchmarkExact time
传统推理框架中每个请求的 KV Cache 需预分配等长连续显存，显存利用率通常仅有 20%-40%，PagedAttention 可将其提升到接近 100%
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
UnverifiedTraditional inference frameworks pre-allocate contiguous KV Cache memory blocks for each request, resulting in an average waste rate as high as 60–80%.83% similarUnverified传统KV Cache分配方案因按最大序列长度预分配连续显存块，导致平均显存利用率仅20%-40%79% similarUnverified推理时的KV缓存、激活值缓冲区和中间计算状态会额外占用30%-50%显存75% similarUnverified传统 LLM 推理中因需按最大长度预留显存，KV Cache 的实际利用率通常不到 30%67% similarUnverifiedThe price of output tokens is typically 4–6 times that of input tokens, stemming from the computational differences between the prefill and decode phases, with GPU utilization during the decode phase potentially being only 10–20% of that during prefill65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/544380API
curl https://kongchang.com/api/v1/knowledge/claims/544380MCP
get_claim(id=544380)