[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"aihot-art-85753":3},{"itemId":4,"vertical":5,"category":6,"source":7,"score":8,"title":9,"summary":10,"analysis":11,"url":12,"coverUrl":13,"direction":13,"marketSignal":13,"publishedAt":14},"85753","ai","论文研究","洪明",67,"Google在Ironwood TPU上优化Qwen 3.5-397B MoE推理性能","Google在Ironwood TPU上优化Qwen 3.5-397B MoE推理，预填充达理论峰值82.4%。","Google 在 Ironwood TPU 上优化 Qwen 3.5-397B MoE 推理性能，探索将大模型推理拆解到硬件边界。\n· 针对 397B 参数的 MoE 模型，在 Ironwood TPU 上进行推理优化。\n· 优化可能涉及模型并行、内存管理、计算调度等硬件级技术。\n· 该工作展示了 Google 在硬件-模型协同优化上的持续投入。\n· 对于超大模型，硬件定制优化是降低推理成本、提升效率的关键路径。\n影响\u002F看点：Google 与 Qwen 的合作表明，硬件与模型联合优化是超大模型落地的必然方向。","https:\u002F\u002Fx.com\u002Fhongming731\u002Fstatus\u002F2077222079985074567",null,"2026-07-15 10:41:51"]