Unverified50% confidenceFactExact time
逐元素操作（如向量加法、激活函数）算术强度仅为0.25-1 FLOP/Byte，属于内存带宽受限操作，因此深度学习框架使用算子融合（kernel fusion）提升算术强度
1
Sources
50%
Confidence
Long-term
Relevance
8/23/2026
First Seen
Sources
Related Entities
Related Claims
Unverified应优先融合访存受限的算子，对算力受限的大型算子（如大矩阵乘法）谨慎融合66% similarUnverified随着MoE架构的普及，激活参数量与总参数量的分离使得FLOP/s阈值失去准确性，DeepSeek-V3训练消耗算力远低于其参数规模所暗示的量级65% similarUnverifiedThe bottleneck of on-device inference is often not peak compute but memory bandwidth; ternary quantization reduces the amount of data that needs to be read per inference by compressing parameters, which is key to achieving 120 tok/s.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/791075API
curl https://kongchang.com/api/v1/knowledge/claims/791075MCP
get_claim(id=791075)