SemiAnalysis · 算力产业· Bryan Shan·· 2026-09-18SelectedAI score76
SemiAnalysis评测Engram架构与DRAM/SSD卸载推理性能
Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading
Not yet translated; shown in Chinese.
AI guide
SemiAnalysis针对Engram模型架构开展推理测试,评估了DeepSeek-V4.1-Flash将多token嵌入表卸载至主机DRAM与SSD的系统表现。
Why it matters
实测展示了将嵌入表卸载至主机DRAM可释放显存并降低并行开销,而SSD卸载路径在当前协同机制下性价比偏低。
Source:SemiAnalysis · 算力产业 · newsletter.semianalysis.com