University of Rostock Institute of Applied Microelectronics and Computer Engineering Monitoring and Control of Temperature in Networks-on- Chip Tim Wegner, Claas Cornelius, Andreas Tockhorn, Dirk Timmermann; MEMICS 2010, Mikulov, Czech Republic, October 22-24

2 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs Outline 1. Introduction 2. Networks-on-Chip (NoCs) 3. Impact of Temperature on Reliability 4. Monitoring & Control of Temperature in NoCs 5. Summary

3 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 1. Introduction  Increasing integration density → rising complexity, shrinking device sizes  NoCs able to deal with arising requirements (e.g. for communication)  But: Reliability becomes a dominant factor for chip design  Goal: Increase reliability in NoC-based systems  Increasing integration density → rising complexity, shrinking device sizes  NoCs able to deal with arising requirements (e.g. for communication)  But: Reliability becomes a dominant factor for chip design  Goal: Increase reliability in NoC-based systems Impacts of technological development

4 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs Outline 1. Introduction 2. Networks-on-Chip (NoCs) 3. Impact of Temperature on Reliability 4. Monitoring & Control of Temperature in NoCs 5. Summary

5 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 2. Networks-on-Chip  Infrastructure for on-chip interconnection  Point-to-point links replace long global busses  Parallel packet-based communication  Separation of communication & computation  Globally asynchronous locally synchronous (GALS)  Modularity of IP cores (not part of actual NoC)  reusability, high abstraction level  Infrastructure for on-chip interconnection  Point-to-point links replace long global busses  Parallel packet-based communication  Separation of communication & computation  Globally asynchronous locally synchronous (GALS)  Modularity of IP cores (not part of actual NoC)  reusability, high abstraction level Properties NoCs are able to satisfy requirements of modern VLSI systems

6 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs Outline 1. Introduction 2. Networks-on-Chip (NoCs) 3. Impact of Temperature on Reliability 4. Monitoring & Control of Temperature in NoCs 5. Summary

7 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 3. Impact of Temperature on Reliability  Increasing integration densities, progress of nanotechnology  Growing number of transistors per chip = raised probability of failure  decreasing structural size of ICs = higher susceptibility to environmental influences & deterioration  Increasing integration densities, progress of nanotechnology  Growing number of transistors per chip = raised probability of failure  decreasing structural size of ICs = higher susceptibility to environmental influences & deterioration Impacts of technological progress Intel 8086 (1978): ≈879 transistors/mm² Intel Bloomfield (2008): ≈2,78 Mio. transistors/mm²

8 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 3. Impact of Temperature on Reliability  Particular physical effects (e.g. TDDB, EM) contribute to deterioration  Abetted by high temperatures  Correlation between temperature & failure mechanisms established by Arrhenius model  Exponential decrease of IC lifetime with temperature  Particular physical effects (e.g. TDDB, EM) contribute to deterioration  Abetted by high temperatures  Correlation between temperature & failure mechanisms established by Arrhenius model  Exponential decrease of IC lifetime with temperature Why is thermal awareness important? Growing influence of on-chip temperature distribution on lifetime, operability, performance etc.

9 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs Outline 1. Introduction 2. Networks-on-Chip (NoCs) 3. Impact of Temperature on Reliability 4. Monitoring & Control of Temperature in NoCs 5. Summary

 Mitigate effects contributing to deterioration & delay occurrence of failures  Control of on-chip temperature distribution  Mitigate effects contributing to deterioration & delay occurrence of failures  Control of on-chip temperature distribution 10 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 4. Monitoring and Control of Temperature for NoCs Objective:  Effective mechanisms to monitor & control on-chip temperature  Integration into existing NoC  Preservation of modularity & reusability  Minimum costs (area, frequency)  Maximum performance of monitoring and control  Minimum impact on system performance  Effective mechanisms to monitor & control on-chip temperature  Integration into existing NoC  Preservation of modularity & reusability  Minimum costs (area, frequency)  Maximum performance of monitoring and control  Minimum impact on system performance Requirements:

11 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 4.1 Mechanisms for monitoring Concept: attach physical monitoring probes to every IP core  temperature variation ∆T  Continuous checking of T IPC  |T IPC,old - T IPC,new | ≥ ∆T ?  Report T IPC,new  Area: 66 LUT/FF pairs  Frequency: 227 MHz  temperature variation ∆T  Continuous checking of T IPC  |T IPC,old - T IPC,new | ≥ ∆T ?  Report T IPC,new  Area: 66 LUT/FF pairs  Frequency: 227 MHz Event-driven:  Period of time ∆t  Report T IPC,new every ∆t  Area: 80 LUT/FF pairs  Frequency: 338 MHz  Period of time ∆t  Report T IPC,new every ∆t  Area: 80 LUT/FF pairs  Frequency: 338 MHz Time-driven:

12 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 4.2 Mechanisms for control  Reception & interpretation of probe packets  Instructions for Dynamic Frequency Scaling to probes (if necessary)  Area: 507 LUT/FF pairs  Frequency: 165 MHz  Reception & interpretation of probe packets  Instructions for Dynamic Frequency Scaling to probes (if necessary)  Area: 507 LUT/FF pairs  Frequency: 165 MHz Central Control Unit (CCU): !!! Not the smartest approach, but suffices to test functionality !!!

 Area penalty: 30,5%  Freq. penalty: 8,2%  Area penalty: 30,5%  Freq. penalty: 8,2%  Area penalty: 7,3%  Freq. penalty: / (but Mux/Demux)  Area penalty: 7,3%  Freq. penalty: / (but Mux/Demux)  Area penalty: /  Freq. penalty: /  Area penalty: /  Freq. penalty: / 13 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 4.3 Integration of monitoring 3 approaches  Different impact on performance & costs Into IP core:Router port of IP core:Extra router port:

14 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 4.4 Impact on system performance

15 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 4.5 Performance of monitoring & control

16 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs 5. Summary  Event-driven approach preferable (situational monitoring, better performance, no redundant traffic, lower area costs)  Integration into NoC using router port of IP core best trade-off between costs & preservation of modularity/non-intrusiveness  Event-driven approach preferable (situational monitoring, better performance, no redundant traffic, lower area costs)  Integration into NoC using router port of IP core best trade-off between costs & preservation of modularity/non-intrusiveness Conclusion Implementation of 2 approaches for monitoring on-chip temperature + 3 methods for integration into NoC  Investigation of: Costs (area, frequency) Impact on system performance Performance of monitoring & control

Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Thanks for your attention! Any questions? University of Rostock, Germany Institute of Applied Microelectronics and Computer Engineering Contact: Homepage:

Establishes relationship between temperature and failure mechanisms Describes dependence of chemical reactions on temperature changes Assumption: all other parameters constant 18 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Arrhenius Model Lifetime of ICs decreases exponentially with temperature Monitoring and Control of Temperature in NoCs

19 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Monitoring and Control of Temperature in NoCs  Inoperability of transistor through gate oxide breakdown (long-term) Time Dependent Dielectric Breakdown (TDDB) Formation of charge traps Current flow !!! HEAT !!! More charge traps Conducting path through gate oxide

20 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Transport of material in conductors (i.e. wires) Cause: ion movement induced by current flow (ions’ mobility increases with temperature) Effects: Hillocks  short circuits Voids  interruption of current paths Electromigration (EM) Monitoring and Control of Temperature in NoCs

21 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Intel Bloomfield: Year: Mio. Transistors 263mm² Tr./mm2 Intel 8086: Year: k transistors 33mm² 879 Tr./mm² Intel Processors Monitoring and Control of Temperature in NoCs

22 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Impact on system performance Monitoring and Control of Temperature in NoCs

23 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Performance of monitoring & control Monitoring and Control of Temperature in NoCs

24 Tim Wegner - 23 October 2010 MEMICS 2010, Mikulov, Czech Republic, October Synthesis results for monitoring & control ComponentIntegration method Event- driven probe Time- driven probe Central Control Unit Into IP core Using IP core port Extra port Frequency [MHz] Area [LUT/FF pairs] Unmodified NoC router: 1771 LUT/FF pairs, 122 MHz Monitoring and Control of Temperature in NoCs