The Future of Electrical I/O for Microprocessors › r5 › denver › sscs › Presentations ›...

1

The Future of Electrical I/O for

Microprocessors

Frank O’Mahony

[email protected]

Intel Labs, Hillsboro, OR USA

2

Outline

• 1TByte/s I/O: motivation and challenges

• Circuit Directions

• Channel Directions

• Tool Directions

• 470Gb/s Prototype

2

3

Microprocessor Bandwidth Needs

• As CPU core count increases, I/O bandwidth (BW) requirements will increase for all segments

• Current system bandwidth requirements (Y2010)– Client BW = ~50GB/s

– Server BW = ~100GB/s

– High-end Server BW = ~200GB/s

Server

Example

High-End

Server Example

4

Microprocessor Bandwidth Trends

Bandwidth Drivers:

CPUMemory

CPUCPU

CPUPeripheral

CPUI/O bridge0.1

1

10

100

1000

Bandw

idth

(G

B/s

)

2000 2005 2010 2015

2X/2yrs.

3X/2yrs.

4

High-end microprocessors are expected to

need ~1TB/s during coming decade

5

Microprocessor I/O Power

• Current system I/O power efficiency is 20-40pJ/bit

• If I/O power efficiency doesn’t improve during the next decade, then:

1TB/s x 20pJ/bit = 160W

System BW I/O Pwr. Eff. I/O Pwr

Client ~50GB/s 20pJ/bit 8W

Server ~100GB/s 20pJ/bit 16W

High-End Server

~200GB/s 20pJ/bit 32W

6

I/O Energy Efficiency Trends

1

10

100

Energy Eff.

(pJ/bit)

2002 2004 2006 2008

-20%/year

Issue: ~20% per year power reduction while bandwidth

increasing 40-70% per year

Year

2010

Ref: R. Palmer,

ISSCC ‘07

7

Energy Efficiency and Channel Loss Tradeoff

• Power efficiency is strongly correlated to channel loss

• Simply scaling per-pin BW will not meet power budget

• Low power interfaces should be “wider” not faster

Based on transceivers

reported 2006-2009 in

65-130nm CMOS

10

100

Energy Eff.

(pJ/bit)

0 10 20 30 401

Channel Loss @ Symbol rate (dB)

7

8

Channel/Interconnect Density

• Conventional package/socket density does not

scale with process

• “Width” of interfaces is limited by routing

congestion

C4 pitch << Pkg. pin pitch

Flip-chip Package

µP

Circuit

area

C4 bump

area

Time

Density

9

Problem Statement Summary

• Bandwidth needs are quickly approaching 1TB/s

• Energy efficiency is not scaling as aggressively

as bandwidth

• The channel limits our ability to increase per-pin

data rate and/or increase the width of an

interface

10

How Will Electrical I/O scale to 1TB/s?

1. Co-design the interconnects and I/O circuitry to

meet bandwidth, scalability and power

efficiency demands

2. Scale the channel by transitioning to new

channel configurations and materials

3. Use accurate, statistical link design tools to

identify balanced architectures.

11

Outline




• Tool Directions


11

12

Low Active Power Techniques

12

10

100

0 10 20 30 401


Pow

er

Eff.

(mW

/Gb

/s)

Power Optimized Links

• Simple equalization

• Low TX swing

• Sensitive RX sampler

• Low-power clocking

13

Minimize analog circuit complexity

RX

L

RTERM

CPAD

Ref: G. Balamurugan, JSSC 4/08

Lowest power links find ways to simplify equalization and

clocking circuitry to reduce power

Equalization examples:

Constrain equalization range by known channel characteristics

Continuous-time linear Rx equalizer

13

14

Power Management: Scalable supplies

• Adapt supply to frequency, process, temperature (f,P,T)

– Digital: Power VSUPPLY2∙f

– Analog: Power VSUPPLY∙Ibias

• Removes excess circuit BW and headroom

VR

VSUPPLY

TX

RX

VREG(f,P,T)

+

-

VREG(f,P,T)

CREG

VREF

VSUPPLY

Regulated supply ring VCO Data link with adaptive supply

14

15

Power Management: Scalable supplies

• Power efficiency improves with adaptive supply/biasing

2.73.6

5.0

11.2

5Gb/s

0.68V

10Gb/s

0.85V

15Gb/s

1.05V

20Gb/s

1.2V

Energy Eff.

(pJ/bit)

0

2

4

6

8

10

12

TX Driver

TX Ser/Pre

TX Clk

RXFE

RX Clk

Refs: G. Balamurugan, JSSC 4/08 and B. Casper, ISSCC ‘06

15

16

Aggressive Power Management

• Don’t spend power doing nothing!

• Rapidly adapt to bandwidth demand

– Requires fast, granular bandwidth adaptation

1.0

0.8

0.6

0.4

0.2

0.0

Normalized

Bandwidth

Demand

Time

17


1.0

0.8

0.6

0.4

0.2

0.0

Normalized

Bandwidth

Demand

Time

Conventional (fixed Bandwidth)

Wasted

Energy




18

Energy

Savings


1.0

0.8

0.6

0.4

0.2

0.0

Normalized

Bandwidth

Demand

Time

Conventional (fixed Bandwidth)

Adaptive

Bandwidth




19

Device Variation in Scaled CMOS

• Device manufacturing

tolerances are improving

• …but area scaling still

causes higher variation

• Fundamental power/area to

variation tradeoff is not

acceptable

Ref: K. Kuhn, IEDM 2007

LeffWeff

c

2

1σV 2

T

19

Need circuit architectures

that fundamentally change

this tradeoff.

20

Mitigating Device Variation

+-Voffset

Circuit derivatives (gm, ro)

are not calibrated by offset

calibration PSRR is not

calibrated

20

• Calibration greatly improves the

power/variation tradeoff

– Receiver offset calibration

– Duty cycle correction

– Adaptive equalizers

– Clock recovery (or deskew)

• Simple calibration doesn’t

alleviate all variation issues (e.g.

PSRR)

21

Mitigating Device Variation• Calibration greatly improves the

power/variation tradeoff

– Receiver offset calibration

– Duty cycle correction

– Adaptive equalizers

– Clock recovery (or deskew)

• Simple calibration doesn’t

alleviate all variation issues (e.g.

PSRR)

• Possible solutions:

– “Dynamic” calibration (e.g. auto-zero)

– Redundancy/reconfigurability

– Better “correct by design” circuits

ΦN

Φ1

selΦ

DSM

ΦOUT

Ref:

P. Hanumolu,

JSSC 2/08.

Coarse

Phase

gen.Φ LPF

DN

L (

ps)

INL (

ps)

21

22

Outline




• Tool Directions


22

23

Channel scaling

• Circuit innovation alone will probably not be enough to

reach the 1TB/s target the channel needs to scale too!

• Better signal integrity: Improved electrical characteristics

mean less power in clocking and equalization

• Higher density: More lanes allow each lane to operate at

lower data rate better power efficiency

23

24

Channel vs. Equalization tradeoffs:

Backplane example

-80dB

-60dB

-40dB

0dB

0GHz 5GHz 15GHz

-20dB

10GHz

5mm Stubbed-via BP

Drilled-via BP

Ref: B. Casper, CICC ‘07.

24

|S21|

25

Improve channel signal integrity

19” Flex

-80dB

-60dB

-40dB

0dB

0GHz 5GHz 15GHz

-20dB

10GHz

5mm Stubbed-via BP

Drilled-via BP

19” Flex Cable

CPU SocketFlex Connector

25

|S21|

26

High density channels

10

1000

Ap

pro

x.

Co

nta

ct/

Ro

utin

g

Pitch

(µ

m)

100

1

Contact pitch

Routing pitch

26

27


10

1000

Ap

pro

x.

Co

nta

ct/

Ro

utin

g

Pitch

(µ

m)

100

1

Contact pitch

Routing pitch

27

28


10

1000

Ap

pro

x.

Co

nta

ct/

Ro

utin

g

Pitch

(µ

m)

100

1

Contact pitch

Routing pitch

28

29


10

1000

Ap

pro

x.

Co

nta

ct/

Ro

utin

g

Pitch

(µ

m)

100

1

Contact pitch

Routing pitch

29

30


10

1000

Ap

pro

x.

Co

nta

ct/

Ro

utin

g

Pitch

(µ

m)

100

1

Contact pitch

Routing pitch

30

31

Outline




• Tool Directions


31

32

What is the “Right” Link Architecture?

• Designers need the ability to quickly and

accurately compare architecture options

32

TX RX

Clock Jitter?

Signal Swing?

Equalization?

ISI? Xtalk?

Modulation (PAM)?

Data Rate?

Interface width?

Clock Jitter?

Sensitivity?

Equalization?

33

Empirical Approach

33

Simulate system with random data

This doesn’t provide adequate accuracy

(BER<10-12)

34

Full System Statistical Analysis

• Specify high-level architecture and block characteristics

• Enables fast evaluation of link sensitivities

Statistical

Signaling

Analysis

TX jitter

Channel & co-channel

responses

EqualizationModulation RX input referred noise

RX sampling jitter

34

35

Maximum Data Rate Comparison:

Backplane vs. Flex

TX FIR taps

DFE taps

1 2 3 4 5 6 4

128

4

1

4

2

4

8

4

16

4

32

4

64

4

4

30Gb/s

15Gb/s

45Gb/s

FlexM

ax. D

ata

Ra

te (

BE

R=

10

-12)

Drilled-via BP

Stubbed-via BP

35

• Statistical system analysis provides designers with real

performance tradeoffs and “brick walls”

36

Maximum Data Rate Comparison:

Backplane vs. Flex

• Statistical system analysis provides designers with real

performance tradeoffs and “brick walls”

TX FIR taps

DFE taps

1 2 3 4 5 6 4

128

4

1

4

2

4

8

4

16

4

32

4

64

4

4

30Gb/s

15Gb/s

45Gb/s

FlexM

ax. D

ata

Ra

te (

BE

R=

10

-12)

Drilled-via BP

Stubbed-via BP

36

37

Outline




• Tool Directions


37

38

txbundle_A[1] rxbundle_B[1]

rxbundle_B[0]txbundle_A[0]

IL-VCOIL-VCO

rxbundle_A[1] txbundle_B[1]

txbundle_B[0]rxbundle_A[0]

rxbundle_A[2] txbundle_B[2]

dieA dieB

fclk_A

9

10

9

9

10

fclk_B

47x10Gb/s, 1.4pJ/bit Interface (45nm CMOS)

38

39

Bundled Architecture

Deskew

Deskew

Deskew

Deskew

Deskew

clk

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

data

Conventional:

Independent clocking

40


• Clocking innovation Bundle clocking

Deskew

Deskew

Deskew

Deskew

Deskew

clk clk

Bundle

Deskew

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

data

Conventional:


41


• Clocking innovation Bundle clocking

Deskew

Deskew

Deskew

Deskew

Deskew

clk clk

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

RX sampler

data data

Bundled clocking reduces I/O power

Conventional:


Optimized:

Bundle clocking

Bundle

Deskew

42

Fast RX Power States

• RX bundle power reduced by 93% in standby

• All RX lanes return to reliable operation in <5ns

0 1 2 3 4 5

0

0.2

0.4

0.6

0.8

1

Wake-up time [ns]

Pro

babili

ty o

f corr

ect

seedin

g

1rxbundle_B[0]

rxlane[9:0]

43

Silicon Area Compression

Conventional: I/O floor plan Power

I/O layout

Ground

I/O signals

44


• Floor plan optimization minimize I/O area


I/O layout

Ground

I/O circuitry

I/O signals

45


• Floor plan optimization minimize I/O area


I/O layout

Ground

I/O signals

I/O circuitry

Optimized: Bundle layout

46

txbundle_A[0]

rxbundle_A[2]

rxbundle_A[1]

rxbundle_A[0]

txbundle_A[1]

IL-VCO + Drv

Lane[9:5] Lane[4:0]

10 TL pairs

Active I/O circuitry

Interface Floorplan1302µ

m

2864µmDie edge

• Active circuit area is reduced with TL routing.

47

Interface Configuration

• Within-bundle lanes matched to <100µm

– Dense LGA connector minimizes breakout area

– Bundles share the same routing layer

– 2X density on stripline layers due to reduced Xtalk

Package

HDI/Flex

bridge500µm LGA

connector

5 signals/mm

Microstrip

Stripline2Stripline1

dieBdieA

SocketPCB

48

Silicon and Interconnect Prototypes

0.5m flex interconnect 3m twinax cable

49

Electrical Interconnect Scaling Challenges

10

100

Power Eff.

(pJ/bit)

0 10 20 30 401


(Based on transceivers reported 2006-2009 in 65-130nm CMOS)

This work: 45nm CMOS

Prototype @10Gb/s

data rate

50

I/O Power Efficiency Measurements

0.00

0.50

1.00

1.50

2.00

2.50

3.00

0 100 200 300

Po

we

r Ef

fici

en

cy (

pJ/

bit

)

Channel Length (cm)

HDI

LCP flex

32AWG -twinax

Link data rate = 10Gb/s*

*high density

interconnect

(HDI)

51

Summary

• Bandwidth needs are quickly approaching 1TB/s

• Extending electrical I/O to 1TB/s requires balance

between power, data rate, density and cost

• Evaluate alternate channel configurations and

materials

• Recent results indicate that electrical will be up to

the task for “in-box” I/O

51

The Future of Electrical I/O for Microprocessors › r5 › denver › sscs › Presentations ›...

Documents

Transcript of The Future of Electrical I/O for Microprocessors › r5 › denver › sscs › Presentations ›...