alan j. drake 1 , xiaobin yuan 2 , pawel owczarczyk 2 , marshall tiner 2

13
Business Unit or Product Name DAC User Track June 6, 2013 Accurate Model-to-Hardware Simulation Methodology for Designing Critical Path Monitors over a Wide Voltage Range Alan J. Drake 1 , Xiaobin Yuan 2 , Pawel Owczarczyk 2 , Marshall Tiner 2 1 IBM Research 2 IBM System and Technology Group

Upload: melva

Post on 23-Jan-2016

37 views

Category:

Documents


0 download

DESCRIPTION

Accurate Model-to-Hardware Simulation Methodology for Designing Critical Path Monitors over a Wide Voltage Range. Alan J. Drake 1 , Xiaobin Yuan 2 , Pawel Owczarczyk 2 , Marshall Tiner 2. 1 IBM Research 2 IBM System and Technology Group. Introduction. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

Business Unit or Product Name

DAC User Track June 6, 2013

Accurate Model-to-Hardware Simulation Methodology for Designing Critical Path Monitors over a Wide Voltage Range

Alan J. Drake1, Xiaobin Yuan2, Pawel Owczarczyk2, Marshall Tiner2

1IBM Research2IBM System and Technology Group

Page 2: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Introduction

DVFS control systems need real-timing timing margin information

Critical path monitors must provide accurate timing measurements across a wide voltage and frequency range– Accurate static timing to check for early and

late mode failures insufficient– High accuracy simulations of extracted netlists

are costly

Developed a method to design a CPM that meets operating ranges while minimizing simulation time.

2

Page 3: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Delay Elements for Fine and Coarse Delay Control

3

Page 4: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Critical Path Monitor Overview

4

Multiple synthesispaths

to mimiccritical paths

Calibration delay

Page 5: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Calibration Delay

Coarse delay block equal to sum of fine delay blocks

Muxes combine delay into continuous range

5

Page 6: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Tuning Methodology

Hierarchical simulation– Fine delay lines only– Synthesis paths without

control

Multiple runs required– 4 paths– 2 guardband

frequencies– 3 operating points:

powersave, nominal, and turbo

– multiple process corners6

Design Data

Extraction

Simulation

Compare to target

=Update designy

n

Page 7: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Extraction and Simulation Hierarchical

– Control: 10724 transistors– Custom synthesis block: 5257 transistors

Extraction– Full coupling for power and noise analysis– Reduced coupling for timing

Simulation– Spice

• Slow (6hrs, 34mins) but flexible for all possible corners• Accuracy dependent on extraction

– Static timing • Fast (50 minutes)• Very accurate at 2 points: early and late mode timing

7

Page 8: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

CPM0 Path1Low-Vt

CPM0 Path2

Low-Vt Wire

CPM0 Path3Mid-Vt

CPM0 Path4

High-Vt

More delay

Less delay

Calibration delay at multiple voltage and temperature points– 4 chips tested– 4 paths per CPM, 5

CPMs per core, up to 8 cores per chip

13 delay step spread

Heavy, steady-state workload

Dark blue: 0.9V, 50C

Magenta: 1.1V, 60C

Cyan: 1.3V, 60C

Yellow: 1.3V, 80C

HW1: Static Timing Only

Target delay

Occ

urr

en

ce

8

Page 9: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013 9

0 3 6 9 12 15 18 210

1000

2000

3000

4000

5000

6000

7000

high-Vt inv mid-Vt invlow-Vt wire low-Vt inv

Calibration Delay Setting:Low Frequency Parts

Occu

rren

ce

Calibration delay at multiple voltages

More than 20 parts tested

High-Vt distribution shifts right and lines up with Low-Vt inverter

Low-Vt distributions shift left

Mid-Vt distribution shifted right

HW2: Static Timing Only After Adjusting for Path Offsets

Page 10: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

HW2: Bit-to-bit Calibration Delay Spreads All Operating Points

10

-228%-162% -95% -28% 38% 105% 172% 238% 305% 372%0

50

100

150

200

250

0to1 1to2 2to3 3to4 4to5 5to6 6to77to8 8to9 9to10 10to11 11to12 12to13 13to1414to15 15to16 16to17 17to18 18to19 19to20 20to2121to22 22to23 23to24 24to25 25to26 26to27 27to2828to29 29to30 30to31

% Change From Nominal

Cou

nts Inverter Coarse

Delay

Interpolating Delay

Page 11: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Simulated Range with More Detailed Extraction:Full Frequency Range

fmax2

0.75 0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.20.5

0.75

1

1.25

1.5

targetpath 1path 2path 3path 4ST

Minimum Insertion Delay

Normalized Voltage

No

rmal

ized

Fre

quen

cy

0.75 0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.20.5

0.75

1

1.25

1.5

targetpath 1path 2path 3path 4ST

Maximum Insertion Delay

Normalized Voltage

No

rmal

ized

Fre

quen

cy

11

Page 12: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Simulated Tuning and Sampling Step Sizes

0.75 V 3s 1 V 3s

Coarse Interface 1.5% 1.3% 1.25% 0.4%

Fine 1.5% 0.5% 1.25% 0.2%

0.75 V 3s 1 V 3s

Bit-to-bit 1.18% 0.3% 1.29% 0.1%

Edge Detector

Tuning delay

• Normalized to nominal cycle time• Includes fine delay and interface

between course step and fine step

• Normalized to nominal cycle time

12

Page 13: Alan J. Drake 1 ,  Xiaobin  Yuan 2 ,  Pawel  Owczarczyk 2 , Marshall Tiner 2

DAC User Track June 6, 2013

Conclusion

Single tool simulation insufficient to guarantee accuracy required over process, corners, and guardband

Hierarchy needed to keep simulation times reasonable

Mix of extraction and simulation environments essential

Accurate capacitive coupling extractions necessary to accurately predict small frequency differences in delay elements

13