Unverified60% confidenceFactTime unknown
The KV cache for a 128K context window of a GPT-4-level model can occupy tens of gigabytes of VRAM
2
Sources
60%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedH100的内存带宽约3.35TB/s，GPT-4规模模型单次生成一个Token需读取约1TB参数数据，理论上每秒最多生成约3个Token79% similarUnverifiedGPT-3 has 175 billion parameters, and storing these weights alone requires hundreds of gigabytes of space76% similarUnverifiedGPT-OSS 20B运行时内存占用约10GB，显存占用约14.5GB75% similarUnverifiedGPU上下文切换需保存完整GPU状态（包括寄存器文件、共享内存和L1缓存），数据量可达数百MB，而CPU上下文切换仅需保存几KB73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/27508API
curl https://kongchang.com/api/v1/knowledge/claims/27508MCP
get_claim(id=27508)