Unverified50% confidenceFactExact time
现代NVIDIA GPU上L1缓存行通常为128字节，全局内存事务的最小粒度为32字节
1
Sources
50%
Confidence
Long-term
Relevance
8/23/2026
First Seen
Sources
Related Entities
Related Claims
Unverified自回归解码在批大小为1时算术强度约等于隐藏层维度（通常4096至16384），而现代GPU峰值算术强度往往超过200，导致解码处于内存带宽受限状态73% similarUnverifiedA single sequence of 32K context length for LLaMA-2 70B can require approximately 80 GB of KV cache in FP16.72% similarVerifiedTaking a 70B parameter model with FP16 precision and 32 attention head layers as an example, the KV Cache for a single 200K-length request requires approximately 80–120 GB of VRAM, far exceeding the 80 GB VRAM limit of a single A100 GPU.71% similarUnverifiedGPU 顶点缓存（Post-Transform Vertex Cache）容量通常仅有 16–32 个顶点槽位70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/791069API
curl https://kongchang.com/api/v1/knowledge/claims/791069MCP
get_claim(id=791069)