Tunable Sensors for Process-Aware Voltage Scaling Tuck-Boon Chan ‡ and Andrew B. Kahng †‡ CSE † and ECE ‡ Departments, UCSD [email protected], [email protected].

1

Tunable Sensors for Process-Aware Voltage Scaling

Tuck-Boon Chan‡ and Andrew B. Kahng†‡

CSE† and ECE‡ Departments, UCSD

[email protected], [email protected]://vlsicad.ucsd.edu

2

Outline• Intro: Adaptive Voltage Scaling (AVS)• Overview of Proposed Method• Voltage Scaling Properties• Designing the Circuit• Results

3

Adaptive Voltage Scaling

• Circuits are designed to guardband for performance variation

• There is margin for typical chips• Adaptive voltage scaling (AVS) adjusts

voltage to reduce power

Voltage

a typical chip

worst-case scenario(e.g, due to process variation)targetf

Maximum frequency

margin

reduce voltage meet performance with less power

Vnominal

4

Taxonomy of AVS Techniques

Open-Loop AVS

Closed-Loop AVSP

ow

er

Freq. & Vdd LUT

Post-silicon characterization

AVS Pre-characterize LUT [Martin02]

Process-aware AVSPost-silicon characterization [Tschanz03]

Generic monitor

Design dependent replica

In-situmonitor

Process and temperature-aware AVS Generic on-chip monitor [Burd00]Design-dependent monitor [Elgebaly07, Drake08, Chan12]

In-situ performance monitor Measure actual critical paths [Hartman06, Fick10]

Error Detection System

Error detection and correction system Vdd scaling until error occurs [Das06,Tschanz10]

Error Tolerance

AVS

approachesAVS classes

5

Motivation for Closed-Loop AVS

• Closed-loop AVS saves up to 62% dynamic power

[Hartman06]

6

Classes of Closed-Loop AVS

• Critical path may be difficult to identify (IP from 3rd party)

• Calibrating monitors at multiple modes/voltages requires long test time

Closed-Loop AVS

Design-dependent replica

In-situmonitor

Generic monitor

• Does not capture design-specific performance variation

This work: Tunable monitor for closed-loop AVS• Can be applied as a generic monitor• Or tuned to capture design-specific performance

7

Outline• Intro: Adaptive Voltage Scaling (AVS)• Overview of Proposed Method• Voltage Scaling Properties• Designing the Circuit• Results

8

Voltage Scaling Key Concepts

• Process distance: process-induced frequency shift relative to target frequency

• Scaling rate: frequency shift (f) per unit voltage difference (V)

• Vmin= Minimum Vdd to meet target frequency• Calculated from process distance and scaling rate

Voltage

SS

k

targetf

nomV)(kVmin_path

Process distance

Max. freq.

Scaling rate =V

f

Monitor Design Concept• Use Vmin of ring-oscillator (RO) as a reference

• Design ROs with worst-case voltage scaling properties an arbitrary circuit will meet target frequency at Vmin_ro

Vmin of ROs Max. Vmin of paths 9

>

V

RO

Critical paths

targetf

Freq.

Process corner A

RO

V

Critical paths

targetf

Freq.

Process corner B

Max.

10

Proposed Method: Tunable Monitor

• Our focus is on voltage scaling property analyze worst-case voltage scaling

Store config.Scenario 2: With chips at process corners

• Extract Fmax and Vmin of chips• Tune voltage scaling properties of ROs

so that Vmin_ro > Vmin_chip

• Recover margin with one calibration

Scenario 1: Without circuit information• Configure RO for worst-case Vmin

• Guardband for arbitrary circuits

Problems• Goal: Vmin_ro > Vmin_path • Questions: Given a process

technology, what is the range of the Vmin that is defined by process distance and scaling rate for arbitrary critical paths?

What circuit techniques can “tune” Vmin?

11

V

Vmin of arbitrary critical paths

targetf

freq.

Vmin

Path BPath A

= ?Also, Vmin changes at different process corners

Path C

12

Outline• Intro: Adaptive Voltage Scaling (AVS)• Overview of Proposed Method• Voltage Scaling Properties• Designing the Circuit• Results

13

Vmin Analytical Derivation

fpath = inverse of average delays of NMOS & PMOS

),(),(),(

nompmosnomnmosnom

2

VkDVkDVkf path

rate scaling

distance processnommin VV

),(),(

),(

nomnom

targetnomnom VkfVVkf

fVkfV

pathpath

path

(1)

(2)

Calculate delays with• Elmore delay model• Effective currents of transistors

(3)

Process distance

Scaling rate

14

Vmin Sensitivity

• Vmin is not very sensitive to fanout, interconnect load, etc.

• Empirically, bounds on Vmin determined by NMOS and PMOS

0.2 0.6 1.0 1.4 1.8 2.2 2.6 3.0 3.4 3.80.9700.9800.9901.0001.0101.020

fanout l beta Ron NMOSRon PMOS

Normalized value of circuit parameters

Nor

mal

ized

Vm

in

wire length

Vmin for PMOS only

Vmin for NMOS only

15

Effects of Fanout and Series Resistance

• Fanout has little effect on Vmin

FO1 FO2 FO4 FO80.600.700.800.901.00

SS TT FF SF FS

Fanout

Vm

in (V

)

1 100 200 400 800 16000.600.700.800.901.00

SS TT FF SF FS

Series resistance (ohm)

Vmin

(V)

High series resistance reduces Vmin

But, need long wires

16

Effects of Cell Type

• Cell type affects Vmin

• Maximum Vmin at different corners are determined by different cell types

• Stacking causes cell delay biased to PMOS or NMOS changes device characteristics and Vmin

INVX0

NAND2X0

NAND3X0

NAND4X0

NOR2X0

NOR3X0

NOR4X00.500.600.700.800.901.001.10

SS TT FF SF FS

Cell type

Vmin

(V)

17

Effects of Cell Strength

• Vmin does not increase from X1 to X3• But increases from X0 to X1•X1 to X3 {1,2,3} fingers, same device characteristic•X0 to X1 Both 1 finger but different diffusion area

X0 X1 X2 X30.500.600.700.800.901.001.10

SS TT FF SF

INV. cell strength

Vmin

(V)

X0 X1 X2 X30.500.600.700.800.901.001.10

SS TT FF SF FS

NAND3 cell strength

Vmin

(V)

Cell layout changes device characteristics and Vmin

18

Outline• Intro: Adaptive Voltage Scaling (AVS)• Overview of Proposed Method• Voltage Scaling Properties• Designing the Circuit• Results

19

Design of RO with Tunable Vmin

• Identified two circuit knobs to tune Vmin

• Series resistance• Cell types (INV, NAND, NOR)

• Proposed circuit• ROs with different cell types (worst-case Vmin are determined

by different cells at different process corners)

• Tune Vmin a configurable series resistance at each stage

1 bit 1 bit 1 bit Control pins

High resistance

Low resistance

20

Tunability• Vmin decreases linearly with % high-resistance

passgates• ROs with different gate types have similar trend

INVX3

0% 12% 24% 36% 48% 61% 73% 85% 100%0.5000.6000.7000.8000.9001.0001.100

SS TT FF SF FS

High resistance passgates (%)

Vmin

(V)

21

Outline• Intro: Adaptive Voltage Scaling (AVS)• Overview of Proposed Method• Voltage Scaling Properties• Designing the Circuit• Results

Experiment Methodology

22

• Goal: Validate PVS ROs in simulation• Check Vmin of ROs vs. Vmin of paths• with arbitrary circuits and process variation

• Experiment setup:• 65nm industrial technology• Implement 3 testcases (arbitrary circuits)• Implement 3 tunable ROs (INV, NAND, NOR)

Power (mW) Area (mm2) Freq. targetFPU 4.1 0.015 710TLU 438.0 0.098 507MUL 19.8 0.050 1042

Process Variation Setup• Simulate critical paths and ROs with SPICE

200 Monte Carlo samples (global variation)• 4 variation sources, Gaussian distributions

• Difference between slow and fast corners define +/- 3 sigma values of variation sources

23

Variation sources mean +/- 3 sigmaNMOS Vth 0 30mVPMOS Vth 0 30mVChannel length 0 5nmGate oxide thickness 0 0.06nm

24

Vmin Extraction and Comparison

• Define ftarget of chip and ROs at “slow-slow” process corner nominal voltage = 1.0V

• Vmin_chip = max. Vmin of critical paths of a testcase

• Vmin_est = max. Vmin of 3 ROs

• For each testcase, calculate Vmin_est - Vmin_chip of every Monte Carlo sample

• A chip is safe when Vmin_est - Vmin_chip > 0

Scenario 1: Guardband for Arbitrary Circuit• Vmin_est - Vmin_chip > 0 under

process variation• Similar results for different

testcases• Small difference between

normal and tunable ROs due to series passgates

25

FPU testcase

MUL testcase

TLU testcase

26

Scenario 2: Tune ROs for Margin Reduction• Extract Vmin_chip at different process corners• Configure % high-resistance passgates

Ensures Vmin_est guided by ROs is always safe

kkVkV ),()( min_chipmin_est

})()({ k

kVkV min_chipmin_estmin. :

s.t. : )},,({max)( cikVkVi

min_romin_est

Experiment Result on Tunability

27

Aggressive config. Vmin_est < Vmin_chip Some chips will fail

Optimized config.• Increase % high

resistance passgates• Vmin_est ≈ Vmin_chip

Default config.

• Low resistance passgates

• Guardband for worst-case

• Vmin_est > Vmin_chip

• 13mV margin

Experiment Result on Tunability

28

Aggressive config. Vmin_est < Vmin_chip Some chips will fail

Default config.

• Low resistance passgates

• Guardband for worst-case

• Vmin_est > Vmin_chip

• 13mV margin

Optimized config.• Increase % high

resistance passgates• Vmin_est ≈ Vmin_chip

Benefits of tunability • Recover voltage margin• Compensate for difference between

SPICE model vs. silicon• Recover margin when chip performance

variation is reduced due to improvements in chip manufacturing

29

Summary• Monitor design based on voltage scaling properties • Estimate the worst-case voltage scaling property

across different process corners Does not require information about critical paths Can be used as an IP for arbitrary circuits

• Recover margin if fmax of sample silicon is available• Future works

Proof of concept silicon Account for performance variation due to layout

context

30

Thank [email protected], [email protected]

http://vlsicad.ucsd.edu

31

Backup Slides

32

Effects of Pass Gates

• Pass gate is equivalent to large resistance• Vmin decreases with fewer parallel pass gates

Vmin decreases

33

Effects of Cell Type and Strength• Key observations: Vmin is affected by cell typesUse NAND, NOR type ROs

• Cell strength changes Vmin Use cells with large Vmin