Zion: Facebook NextGeneration Large Memory Training Platform
Zion平台作为AI训练系统,主要分为8插槽服务器、8加速器平台和OCP加速器模块三个主要部分。同时,它采用了Facebook的OAM模块(OCP Accelerator Module),在设计上还能够处理一系列神经网络,包括CNN、LSTM和SparseNN等。
Conclusions
• Zion is FB next generation flexible training platform • Co-designed to target demanding recommendation models • Adopts new vendor-agnostic OCP accelerator module • Building block that can scale out to a bigger system