Unverified70% confidenceFactExact time
GPU存储层次从快到慢为：寄存器、L1缓存/共享内存、L2缓存、全局显存
2
Sources
70%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
UnverifiedGPU多级内存层级从高到低依次为：寄存器文件（延迟约1周期）、共享内存/L1缓存（约32-228KB，延迟约20-40周期）、L2缓存（H100约50MB，延迟约200周期）、HBM全局内存（延迟约500-800周期）85% 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KB82% similarUnverifiedGPU的SM能同时驻留的Warp数量决定了GPU隐藏内存访问延迟的能力，制约Occupancy的核心资源是寄存器文件与共享内存79% similarUnverifiedGPU Direct Storage（GDS）通过建立NVMe SSD与GPU显存之间的直接DMA通道，绕过CPU内存，在实际测试中可将大文件加载带宽提升2-5倍78% similarUnverifiedA new storage-inspired memory solution could enable a single GPU's usable memory capacity to grow to the multi-terabyte level77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/511249API
curl https://kongchang.com/api/v1/knowledge/claims/511249MCP
get_claim(id=511249)