Regular Session I

on YouTube ↗01h 05m

" summary

Three concrete deadlock solutions and measurements: an opportunistic traversal packet bypass that tags a “travel” flit to escape a dependency loop, implemented with extra control signals and evaluated on a gem5 simulator, reporting lower average latency and higher saturation throughput versus prior recovery schemes. A supermesh + DOR + bubble-flow-control recipe implemented in SystemC enforces four routing rules and cuts worst-case latency by up to 35% under heavy broadcast traffic while BFC can raise average latency ~14% unless buffers grow. JARVA encodes SMT-based joint routing and static VC allocation to minimize max channel load, validated with Noxim/HORNET and showing higher injection-rate saturation and lower latency than standard oblivious/adaptive baselines.

AI-written summary · about

" chapters (20)

1. 0:00Session introduction
2. 2:51Deadlock recovery in NoC
3. 10:10Traffic patterns motivation
4. 12:18Bypass mechanism hardware
5. 15:54Experimental evaluation results
6. 18:04Contributions future directions
7. 19:11Q&A deadlock recovery
8. 23:14Deadlock freedom in simulators
9. 25:20Neuromorphic system challenges
10. 29:30Deadlock avoidance baseline
11. 32:25Deadlock-free routing rules
12. 36:19Dimension order routing control
13. 38:06Conclusion future research
14. 38:28Q&A deadlock freedom
15. 45:59JARVA routing framework
16. 48:09Deadlock avoidance problem
17. 53:06Routing virtual channel allocation
18. 58:04SMT solver approach
19. 1:01:35Validation performance results
20. 1:02:19Q&A JARVA wrap-up