1 / 26

Roy Lee Advisor: Lei He royjylee@ucla eda.ee.ucla October 26, 2011

SEU Mitigation for FPGA- based Systems. Roy Lee Advisor: Lei He royjylee@ucla.edu http://eda.ee.ucla.edu October 26, 2011. 1. Outline. Introduction In-Place Decomposition (IPD) In-Place LUT Polarity Inversion (IPV) Experimental Results Conclusions & Future Works. Robustness in FPGAs.

bevan
Télécharger la présentation

Roy Lee Advisor: Lei He royjylee@ucla eda.ee.ucla October 26, 2011

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. SEUMitigation for FPGA-basedSystems Roy Lee Advisor: Lei He royjylee@ucla.edu http://eda.ee.ucla.edu October 26, 2011 1

  2. Outline Introduction In-Place Decomposition (IPD) In-Place LUT Polarity Inversion (IPV) Experimental Results Conclusions & Future Works

  3. Robustness in FPGAs FPGAs are extensively used not only for prototyping but also in a wide range of applications such as internet networking and communication equipment, and robustness is among the most important design objectives An effective approach for reducing the impact of Single Event Upset can lead to higher mean-time-to-failure(MTTF), increased quality of service, and reduced maintenance cost

  4. Single Event Upset (SEU) Single Event Upsets (SEUs) : one of the main causes of reliability reduction caused by charge particle strikes due to cosmic radiation which create soft errors Major effect on circuits : change the logic state of a static memory element Trend : SEU vulnerability is increasing with technology shrinking

  5. SEU in FPGAs Most of the commercial FPGAs employ SRAM as their configuration memory elements for higher logic density and programming flexibility Three of the major memory elements in FPGAs : user flip-flop, block RAM, and configuration RAM

  6. Single Event Upset in FPGA The circuit effect of SEUs in a FPGA is permanent until the FPGA is re-programmed • Interconnect : • Logic : SEU on Configuration RAM is much more critical!

  7. Demand for In-Place Reliability Optimizations • Triple Modular Redundancy (TMR) is the most popular fault tolerant technique, but it requires more than 3X overhead on power, area, and cost • For non-mission critical applications, such as communication systems, robustness improvement with little or no overhead is highly demanded • In-place optimization techniques provide reliability improvement while preserving circuit placement and routing, and therefore the overhead is minimal

  8. In-Place Resyntheses Flow Mitigation after placement and routing without change of placement and routing (and no change on design closure) Design Entry In-Place Resyntheses Logic Synthesis In-Place Decomposition (IPD) Map In-Place LUT Polarity Inversion (IPV) Placement and Routing Bitstream

  9. Outline Introduction In-Place Decomposition (IPD) In-Place LUT Polarity Inversion (IPV) Experimental Results Conclusions & Future Works

  10. Fault Metrics • Soft Error Rate(SER) of a configuration SRAM bit: • Mean-Time-To-Failure (MTTF) • System level measurement of reliability • For single fault model, MTTF  1/average(SERb)

  11. In-Place LUT Decomposition Leveraging the dual-output feature of LUT architecture and the built-in carry chains Dual-output 6LUT • Xilinx Virtex-5 6-input LUT architecture

  12. LUT Decomposition • Decomposition : F = C( F1, F2, ……, Fn ) (C is the converging logic function) Decomposition Original LUT Decomposed LUT

  13. Example 1 : In-Place Duplication The average SER of the LUT is : (S0+……+ S31)/32 …… …… …… 5-input 5-input AND function

  14. The average SER of the LUT is reduced to (2*S31)/32 Example 1 : In-Place Duplication 0 -> 1 Covered …… …… …… 0 Covered …… …… ……

  15. The number of SRAM bits used is reduced from 32 to 12, and the SERs of unused bits are 0 The average SER is also reduced due to logic masking of the converging logic Example 2 : In-Place Decomposition …… …… ……

  16. Outline Introduction In-Place Decomposition (IPD) In-Place LUT Polarity Inversion (IPV) Experimental Results Conclusions & Future Works

  17. Fault Masking for MUX • Fault is masked when logic(i) = logic(j) SEU on a routing MUX

  18. Exampleof Fault Masking SER(bk)=( v(i) v(j) ) · observ(m) observ(m) is the fault observability at MUX m : the probability of the fault that can be propagated to the primary outputs

  19. LUT Polarity Inversion LUT inversion Fanout adjustment Polarity can be determined independently for each input and the output of an LUT

  20. Inversion to Reduce SER SER: 1-0.9*0.2-0.1*0.8=0.74 SER: 1-0.9*0.8-0.1*0.2=0.26

  21. Outline Introduction In-Place Decomposition (IPD) In-Place LUT Polarity Inversion (IPV) Experimental Results Conclusions & Future Works

  22. Improvement by IPD SER reduction for MCNC benchmarks mapped to 6-input LUTs IPV increase LUT-level MTTF by 4.52x, and chip-level MTTF by 1.07x (due to dominance of interconnects)

  23. Improvement by IPV SER for MCNC benchmarks mapped to 6-input LUTs IPV on average increases chip-level MTTF by 3.07X Less than 50% LUTs need to be inverted

  24. Improvement by Combined Algorithms Averaged SER reduction for MCNC benchmarks mapped to 6-input LUTs IPF+IPD+IPV reduces chip-level SER by 70.53%  3.39x chip-level MTTF increase ZheFeng, Naifeng Jing and Lei He,“IPF: In-Place X-Filling to Migrate Soft Errors in SRAM-Based FPGAS,” FPL 2011

  25. Conclusions & Future Works Proposed two robust resynthesis techniques, In-Place Decomposition(IPD) for logic and In-Place LUT Polarity Inversion(IPV) for interconnect, to improve circuit robustness without global overhead We show on average 3.39X MTTF improvement on the MCNC benchmark circuits when combining IPD, IPV, and IPF In the future, we will develop more in-place resynthesis techniques and investigate the interaction among different techniques

  26. Thank you!

More Related