In this chapter, we present ClosNN, a specialized NoC for NNs based on the well-known Clos topology. Clos is perhaps the most popular Multistage Interconnection Network (MIN) topology. Clos is used commonly as a base of switching infrastructures in various commercial telecommunication and network routers and switches.

Chapter Contents:

• 3.1 Related work
• 3.2 Preliminaries
• 3.3 ClosNN: a customized Clos for neural network
• 3.4 Collective communications on ClosNN
• 3.5 ClosNN customization and area reduction
• 3.6 Folded ClosNN
• 3.7 Leaf switch optimization
• 3.8 Scaling to larger NoCs
• 3.9 Evaluation
• 3.9.1 Performance comparison under synthetic traffic
• 3.9.2 Performance evaluation under realistic workloads
• 3.9.3 Power comparison
• 3.9.4 Sensitivity to neural network size
• 3.10 Conclusion
• References

Inspec keywords: parallel architectures; network-on-chip; feedforward neural nets

Other keywords: ClosNN; network routers; feedforward neural network; massively parallel architectures; multistage interconnection network topology; switching infrastructures; NoC; switches; MIN topology

Subjects: Neural computing techniques; Network-on-chip; Parallel architecture; Network-on-chip