Unverified50% confidenceBenchmarkExact time
PagedAttention 理论上可将显存利用率从不足 40% 提升至 90% 以上
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified传统LLM推理中KV缓存的实际利用率往往不足50%,PagedAttention可提升至接近100%75% similarUnverifiedPagedAttention可将GPU显存利用率从通常的20%-40%提升至接近理论上限,吞吐量相比原生HuggingFace推理提升数倍至十余倍73% similarUnverified发帖者指出编码任务中缓存命中率通常高达90%以上72% similarUnverifiedTailscale官方数据显示直连成功率超过90%65% similarUnverified在提升召回率时不可避免地牺牲精确率,将召回率从95%提升至99%可能使精确率从90%降至70%65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/898236API
curl https://kongchang.com/api/v1/knowledge/claims/898236MCP
get_claim(id=898236)