Csaba Andras Moritz © 2007 Microprocessor Design in the Face of Process Variations Csaba Andras Moritz Electrical & Computer Engineering University of.

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Outline Introduction Impact of Process Variations A Process Variation Resilient Pipeline A Process Variation Resilient Adaptive Cache Architecture Results Conclusion

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Introduction As technology scales, the feature size reduces thereby requiring a sophisticated fabrication process. The process variations increase as the feature reduces due to the difficulty of fabricating small structures consistently across a die or a wafer. These variations cause mismatches between identical structures. With respect to circuits, this translates to a change in all devices or interconnects parameters from their mean value. Device and interconnect variation trends for different technology generations Sani Nassif, etl. “Models of Process Variations in Device and Interconnect”. IEEE Press 2000

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Two main sources of process variation:  Physical factors (intrinsic variation)  Environmental factors (dynamic variation) The physical factors are permanent and result from limitations in the fabrication process  Effective Channel Length (Geometric Variations): Imperfections in photolithography (mask, lens, photo system deviations)  Threshold Voltage (Electrical Parameter Variation): Variation in device geometry Random dopant fluctuations changes in oxide thickness The environmental factors depend on the operation of the circuit and include variations in:  Temperature, Power Supply, Switching Activity The performance and power consumption of integrated circuits can be greatly affected. Introduction

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 15 NAND gates and NAND2 C load “1”“1” “1”“1” “1”“1” A = “ 1 ” B = “ 0 ” → “ 1 ” C = “ 1 ” → “ 0 ” 15 NAND Gates A B C V BN V BP V BN

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Assumptions We target a future 32-nm technology process where leakage and process variation are significant In the nominal delay we assume there is no process variations impact on the pipeline stage. In worst-case we assume the worst values of the parameter variations at each transistor that will result in the maximum delay or power consumption. A body bias is a voltage applied between the source or drain of a transistor and its substrate, effectively changing the transistor’s Vth. Depending on the polarity of the voltage applied, Vth increases or decreases. If it increases, the transistor becomes less leaky and slower (reverse body bias); if it decreases, the transistor becomes leakier and faster (forward body bias). Table 1 shows parameter values of process variations for different cases. Figure 3 and Table 2 show delay of the pipeline at different body bias voltages. Figure 4 and Table 3 show average power consumption of the pipeline stage with different body bias voltages.

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Device parameter variations Leff, Vdd, and Vth Table 1. Parameter values for different cases Case Effective Channel Length (L eff ) Supply Voltage (V dd ) Threshold Voltage (V thn )(V thp ) Nominal25.32 nm0.90V0.20V-0.21V Best-case20.26 nm0.96V0.18V-0.19V Worst- case 30.38 nm0.84V0.22V-0.23V

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Table 2. Delay of the pipeline stage. Case Nominal Body Bias Delay V BN V BP Nominal0V0.9V1.363 ns Best- case 0V0.9V0.646 ns Worst- case 0V0.9V3.811 ns Case Forward Body Bias Delay V BN V BP Nominal0.5V0.4V1.271 ns Best- case 0.5V0.4V0.631 ns Worst- case 0.5V0.4V3.389 ns Case Reverse Body Bias Delay V BN Nominal-0.5V1.4V1.608 ns Best- case -0.5V1.4V0.696 ns Worst- case -0.5V1.4V4.731 ns Delay of Pipeline Stage

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Power of Pipe Stage Table 3. Average power of the pipeline stage. Case Nominal Body Bias Average Power V BN V BP Best- case 0V0.9V7.843 μW Nominal0V0.9V22.45 μW Worst- case 0V0.9V219.4 μW Case Forward Body Bias Average Power V BN V BP Best- case 0.5V0.4V13.00 μW Nominal0.5V0.4V30.32 μW Worst- case 0.5V0.4V294.5 μW Case Reverse Body Bias Average Power V BN Best- case -0.5V1.4V7.772 μW Nominal-0.5V1.4V19.68 μW Worst- case -0.5V1.4V178.7 μW

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Effect of BB on delay and power Table 4. Effect of Body Bias Technique. Case Body Bias Voltages Delay (ns) Average Power (μW) V BN V BP Forward Body Bias 0.85V -0.6V 1.087677.0 0.65V -0.1V 1.271410.9 0.50V 0.4V 1.27530.32 Nominal0V 0.9V 1.36322.45 Reverse Body Bias -0.5V 1.4V 1.60819.68 -1.0V 1.9V 1.94117.59 -1.5V 2.3V 2.34616.94

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 All parameters summary Table 5. Effect of all parameters on pipeline delay Maximum (ns) 1.703 Minimum (ns)1.214 Mean (ns)1.389 Sigma0.056

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Summary power consumption Table 6. Effect of all parameters on pipeline power consumptions. Maximum (uW) 29.65 Minimum (uW)19.51 Mean (uW)24.05 Sigma1.168

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Razor Latches Latch concept to sample output of a stage two different times Compare outputs If not equal resample inter-stage latch and delay pipeline by one cycle Implications?

Recovery Technique 1: Global Clock Gating If any stage detects a timing problem  Stall the entire pipeline for one clock cycle.  Use this additional clock cycle to recompute using the correct shadow-latch values

Recovery Technique 2: Counterflow Pipelining When a mismatch (between regular and shadow latch contents) is detected:  Assert a bubble signal, to specify that the erring pipeline slot is now to be considered a bubble.  In the subsequent cycle, inject the shadow latch value into the next stage, allowing the errant operation to continue with the correct values  Trigger a flush train, traveling backwards from the errant stage, flushing operations at each stage it visits

Process Variation Impact on Memory Systems The process variations are random in nature and are expected to become significant in the smaller geometry transistors commonly used in memories. Process variations in caches affect the performance of circuits like  Sense amplifiers that require identical device characteristics  SRAM cells that require near-minimum-sized cell stability for large arrays in embedded, low-power applications  The delay of the address decoders suffer from the process variations that can result in shorter time left for accessing the SRAM cells Question is whether there is a significant delay variation overall that will drive a change in memory architecture design.

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 To account for the worst-case scenario we might need to increase the cache access time by 2 to 3 cycles in conventional design. Application performance could be impacted by as much as 30-40%! These results suggest that process variations must be taken into consideration  New types of circuits and architectures? Motivation

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 There are several ideas that could be exploited in a memory system:  reduce performance by operating at a lower clock frequency (conservative approach)  increase cache access latency assuming worst-case delay (conservative approach)  variable-delay cache architecture (adaptive approach) Introduction

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 The focus of this presentation is on CAM-based caches. Cache Organization Overview Matchline CAM Tags Data SRAM Cache Bank 32 lines 8 words 16 Banks Tag Bank Word Byte 31 9 8 5 4 2 1 0 Virtual Address: MUX Data

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Cadence tool was used to design the circuits at layout level, and HSPICE simulation used to evaluate the performance. All the circuits were designed using 32-nm CMOS technology and simulated with a supply voltage of 0.9V. Experiment Setup Cache ComponentPower Techniques Bank Decoder4-input Static NOR gates Tag Array10-transistor CAM Cell Data Array6T SRAM Cell Cache lineWordline Gating Line decoderTwo level decoding: 1 st level 3-input DNAND gate and 2 nd level 2-input NOR gate Tag & Data ArraysCache subbanking (16 banks) Bank size1KB Sense AmplifiersAlpha latch & Sharing Sense Amps. Configuration of our 16 KB Low Power Cache

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Effective Channel Length variation:  Imperfections in photolithography (mask, lens, photo system deviations) A 40% variation in Leff is expected within a die [ Sani Nassif, IEEE press 2000 ]. Worst-Case Conditions

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Effective Channel Length variation: A small variation in the Leff value causes a change in the leakage power by as such as 60X from the nominal value. Worst-Case Conditions

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Threshold Voltage Variation:  Accurate control of Vth is very important for many performance and power optimizations and for correct execution. Worst-Case Conditions

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Threshold Voltage Variation  The impact on leakage power could be as much as 40X. Worst-Case Conditions

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Power Supply Variation  One of the most important environmental factors that cause variations in operating condition is supply voltage.  Voltage variations due to non uniform power-supply distribution, switching activity, and IR drop;  A total variation of 15% in Vdd was considered with a nominal value of 0.9V. Worst-Case Conditions Vdd (V)Delay (ns)Power (W) 0.830.7460.183 0.860.7170.187 0.900.6670.191 0.930.6340.213 0.970.6010.266

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 To accurately predict cache critical path delay distribution at the circuit level, cache delay variability can be studied through Monte-Carlo in HSPICE circuit simulations. Monte-Carlo simulations verify model predictions over a wide range of process and design conditions and provides an estimate for expected behavior. We assume parameter variations to be normally distributed with mean and sigma values derived from PTM and ITRS sources. Expected Conditions Technology32nm DeviceNMOSPMOS Leff25.32 nm (± 20%) Vth0.2V (± 7.5%)-0.2V (± 7.5%) Vdd0.9V (± 7.5%) Temperature75 o C Parameter values and σ variations

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 The distribution of delay of a cache critical path was determined by performing Monte-Carlo sampling at different supply voltages, threshold voltages, and transistor lengths. under the expected condition a large fraction of accesses would be still close to the nominal value Nominal Expected Conditions

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 How do we design a memory system in the face of process variations and help mitigate the negative impact on performance? We can select a cache design using worst case assumptions  ALL VARIATIONS and ALL COMPONENTS on the critical path Alternatively, we need to design circuits and architectures that would work adaptively depending on actual delay  Process variation resilient design  Resilience against delays in different parts of the cache Architectural Techniques

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Two phases of operation: classification and execution Proposed Adaptive Cache Architecture Data Array Data Array CAM Tag CAM Tag FDEXMEMWB Delay Storage Delay Storage Adaptive Controller Test Mode Classifier addressdata

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 During classification phase  The cache is equipped with a built-in-self-test (BIST) technique to detect speed difference due to process variation.  Each cache line is tested using BIST when the test mode signal is on. A block is considered medium, slow, failure. Delay Storage Column MUX Sense Amplifiers Data Out Row Address Data Array BIST Test Mode Speed Information Operating Conditions Classification Phase

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Delay Storage Controller Column MUX Sense Amplifiers Data Out Column Address Row Address Data Array During execution phase  The speed information stored in the delay storage is used to control sense amplifiers during regular operations of the circuit. Execution Phase

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Instruction WindowRUU=16; LSQ=8 Fetch, dispatch, commit width4 Integer ALU/multi-div4/1 FP ALU/multi-div4/1 Number of Banks16 banks L1 D-cache Size16KB, 32-way set-assoc, 32B blocks L1 I-cache Size16KB, 32-way set-assoc, 32B blocks L2 Unified Cache Size128KB, 64-way, 64B blocks, 8cycle Memory Latency100 cycles Memory ports2 TLB Size128-entry, fully assoc., 30 cycles miss penalty Branch PredictorComb. Of bimodal and 2-level gshare; bimodal size 2048; level 1 1024-entry, history 10; level 2 4096- entry (global) Branch Target Buffer512-entry, 4-way associative Return-address-stack8-entry Experimental Setup The adaptive cache architecture is implemented in the SimpleScalar. We have conducted simulations of SPEC2000 benchmarks using the adaptive approach. The adaptive cache based on the delay distribution is determined by the Monte-Carlo simulation. SimpleScalar parameters for CPU

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Baseline: 3 cycle D-cache with worst-case delay, 16KB total size, 16 banks each 32-way. Out of order 4-way issue. Adaptive caching scheme: 1% 3 cycle, 24% 2 cycle. 75% 1 cycle cache line access. Results below show performance is improved by 9% to 31%! Performance Speedup

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Speedup values are normalized with respect to the worst-case delay of 3 cycles. As we can see, the 8-way issues design benefits more than the 4-way issues from the adaptive cache architecture. Sensitivity to Issue Width

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Hardware required :  BIST circuit  delay storage  control circuitry We have evaluated the hardware needed for the adaptive cache by using the Synopsys Design Compiler tool. Hardware Required CircuitBIST, delay storage, and control circuitryCache Delay0 ns0.95 ns Power0.55 mW27.67 mW Area0.0048 mm^20.54 mm^2

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Leakage Enhanced Cells In the inactive state, when the cell is not being written to or read from, most of the leakage power is dissipated by the transistors that are off and that have a voltage differential across their drain and source. If the cell were storing a “0”, transistors T1, N1 and P2 would dissipate leakage power. A simple technique for reducing leakage power would be to replace all transistors with high-Vth ones, but this would degrade the bitlines discharge times affecting cell read performance significantly. In our design we instead applied the same high-Vth for all the NMOS transistors – asymmetric cell design. By changing the Vth we change perfomance and power tradeoffs.

Csaba Andras Moritz - Software Systems & Architecture Lab, Electrical & Computer Engineering; © 2007 Tradeoffs between performance and power – what is visible at appl. level? Distribution of cache delay and leakage power for different high-V th schemes. Results obtained by Monte Carlo simulations with adaptive cache for various scenarios. Scheme V th (V) Delay (ns) Mean Leakage (W) 1 cycle 2 cycles 3 cycles Conventional0.232.340.1900% 100% A10.200.9520.46775%24%1% A20.250.9720.18268%30%2% A30.271.0910.11656%40%4% A40.301.1220.07645%50%5%

Csaba Andras Moritz © 2007 Microprocessor Design in the Face of Process Variations Csaba Andras Moritz Electrical & Computer Engineering University of.

Similar presentations

Presentation on theme: "Csaba Andras Moritz © 2007 Microprocessor Design in the Face of Process Variations Csaba Andras Moritz Electrical & Computer Engineering University of."— Presentation transcript:

Similar presentations

About project

Feedback

Log in

Auth with social network:

Csaba Andras Moritz © 2007 Microprocessor Design in the Face of Process Variations Csaba Andras Moritz Electrical & Computer Engineering University of.

Similar presentations

Presentation on theme: "Csaba Andras Moritz © 2007 Microprocessor Design in the Face of Process Variations Csaba Andras Moritz Electrical & Computer Engineering University of."— Presentation transcript:

Similar presentations

About project

Feedback