Unverified50% confidenceFactExact time
100万Token在FP16精度下仅KV Cache一项即可消耗数百GB显存，远超单张A100/H100的80GB容量上限
1
Sources
50%
Confidence
Long-term
Relevance
7/22/2026
First Seen
Sources
Related Claims
Unverified支持128K上下文的70B参数模型的KV Cache满负荷时可能需要数百GB显存，超过单张H100的80GB显存容量85% similarUnverified以LLaMA-2 70B为例，处理4096 token上下文、批量大小32时，KV Cache可消耗超过80GB显存82% similarUnverified对于拥有32层、32头注意力的典型7B模型，每个token需约0.5MB显存存储KV Cache，生成1000个token则需额外约500MB显存76% similarUnverified百万Token级别的单次推理需要数十至数百GB的高端GPU显存（A100/H100级别）76% similarUnverified搭载AMD 780M/Radeon 890M核显的迷你机，需插满双通道内存（2x16GB以上）才能发挥完整核显性能，单条内存会导致核显帧率腰斩约40%76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/593698API
curl https://kongchang.com/api/v1/knowledge/claims/593698MCP
get_claim(id=593698)