plik

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 2 September 1999

&DUQHJLH 0HOORQ

CHIPS Team

• Center for Highly Integrated Information

Processing and Storage Systems (CHI

• Current Research Team:

– L. Richard Carley, James A. Bain, Gary K. Fedder, Greg Ganger,

John Griffin, David W. Greve, David F. Guillou, Michael S. C. Lu,
Tamal Mukherjee, David Nagle, Suresh Santhanam, Steve Schlosser
Dept. of Electrical and Computer Engineering, Carnegie Mellon University

Leon Abelmann,
Information Storage Technology Group, MESA Research Institute,
University of Twente, Enschede, The Netherlands

Seungook Min
Dept. of Materials Science and Engineering, Carnegie Mellon University

• Support:

– DARPA ETO - Since 1993

– Industry: Intel, Texas Instruments, Infineon, Storage Tek

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 3 September 1999

&DUQHJLH 0HOORQ

• Highly Integrated Info. Processing and Storage Systems

–

Processor (> 500 MIPS)

–

RAM (> 64 MB)

–

Nonvolatile Mass Memory (> 1 GB)

–

Communications (> 100 MB/s)

• IC Processing Based

–

Manufactured by IC
photolithographic techniques
(parallel manufacturing)

• Advantages of Integration

–

Less expensive

–

Smaller volume and mass

–

Lower power

–

Co-location of Storage and Processing

archival data storage / secure data

–

High shock resistance possible

–

NEW APPLICATIONS and NEW MARKETS

The Vision

2 cm

CPU

RAM

Communications

Nonvolatile

Mass Memory

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 4 September 1999

&DUQHJLH 0HOORQ

EEPROM for Mass Storage ??

• We have computers on a chip now - Embedded computers

– Billions of embedded CPUs sold today (vs. 100’s of millions of PC CPUs)

– How are HI

systems different from “embedded computer” of today??

» Currently nonvolatile memory is EEPROM (FLASH memory)

» Dramatic increase in amount of nonvolatile mass memory (many GB)

» Quantum leap in kinds of applications possible; e.g., (1) a powerful OS!

(2) large data space, (3) complex programs, (4) learning (store history)

⇒

INTELLIGENT BEHAVIOR

• EEPROM Scaling vs. Time (

high voltage

⇒

hard to scale

– From Semiconductor Industries Association (SIA) Roadmap 1997/98

EEPROM 1997 1999 2002 2005

2008 2011

– NOR Cell Area 0.6µm

0.3µm

0.15µm

0.08µm

0.04µm

0.02µm

(density MB/cm

) 16 32 64 120

240 480

– EEPROM cost $/MB 6 3 1.5 0.8

0.4 0.2

(Best Case - no increase in fab cost / cm

)

– Expect IC-based mass storage at 10 GB/cm

by 2008

(>40X)

• Taking EEPROM prices as $0.4/MB

⇒

10 GB = $4,000 !!

– For IC-Based mass storage in 2008 predict < $40 / 10 GB

(> 100X better)

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 5 September 1999

&DUQHJLH 0HOORQ

Motive: Lower Entry Cost

• Motivation:

– Possible new product niches

– Divisible cost

facilitates many applications

– Can be merged with DRAM & CPU(s)

• Goal:

– Complete computing & storage system

– $20 - $30

– 1 or a few IC substrates

• Example Applications:

– “throw-away” sensors / data logging systems

  infrastructure monitoring; e.g., bridge monitors, concrete pours,
  smart highways, condition-based maintenance, security systems,
  low-cost speaker-independent continuous speech recognition, etc.

– “ubiquitous” use in everyday world

every appliance will be “smart” and “communicative”

– NOTE: embedded CPUs common, but no room for sophisticated code

» Intelligent behavior requires LOTS of SOFTWARE

0.01 GB

0.1 GB

1 GB

10 GB

100 GB

$10

$100

$1000

CACHE RAM

DRAM

HARD

DISK

Entry Cost

Capacity @ Entry Cost (est. 2005)

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 6 September 1999

&DUQHJLH 0HOORQ

Motive: Low Volume & Mass

• 10 GB/cm

= 65 GB/in

density (100x CD-ROM)

• Expect stored bit size 20 nm x 20 nm bit size by 2009

• Example Applications:

» Space / satellite use - store data when not in line of site

act as packet buffer for communications satellites, etc.

» Human portable applications - e.g., medical implants, super PDA

» Law enforcement / monitoring devices / security surveillance

0.1

100

1000

10,000

0.1

100

1000

10,000

100,000

Storage Capacity [GB]

3.5” Disk Drive

Flash memory, 0.1 µm

/ cell (no packaging)

Occupied
volume [cm

]

@ 10 GB/cm

(no packaging)

1” Disk Drive

Prediction

2005

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 7 September 1999

&DUQHJLH 0HOORQ

Motive: Storage System on a Chip

• Completely Integrated System

⇒

single chip = processing + mass storage

–

Enables new level of data archiving - push “lifetime” of data

Data reading hardware is archived with data

Software that can interpret data can be archived with data

Hardware that executes that software can be archived with data

Final interface can be simple long-lived standard one (e.g., RS-232)

–

Enables new level of “secure” devices

Internal IC lines extremely difficult to access

2 cm

USER DATA

Applic. Software

Hardware (CPU)

Generic Comm.

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 8 September 1999

&DUQHJLH 0HOORQ

Motive: Lower Data Latency

• Motivation:

– Many types of software limited by access time to data

» Generally when future accesses hard to predict

(e.g., Transactions Based Processing)

– Conventional disk drives limited by rotational velocity:

» at 10,000 RPM worst case rotational latency is about 6ms

» at 30,000 RPM (dental drill speeds!!) still 2ms

– IC-Based Mass Store: latency designable param. - 100’s of µs possible

Worst-Case

Access

Time

Cost $ / GB

$1 / GB

$10 / GB

$100 / GB

$1000 / GB

10ns

1µs

100µs

10ms

DRAM (volatile)

HARD DISK

Prediction

2005

EEPROM (Flash)

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 9 September 1999

&DUQHJLH 0HOORQ

Building the Hardware

• Integration already achieved for CPU, RAM, Comm.

• The Major Challenge: Si-Based Nonvolatile Mass Memory

–

Manufacturing nonvolatile mass memory systems using IC wafer fab

Compatibility with IC limits processing temperatures and chemistry

• Existing Approaches to building Si integrated NV RAM

–

Today’s approaches are all based on lithographically defined cells

–

For Example: EEPROM (flash memory) - Block erasable NV RAM
FE-RAM and M-RAM new approaches - direct rewrite NVRAM

–

Whats the problem?: Lithography!! Typical storage cell > 3

λλλλ

x 3

λλλλ

= 9

λλλλ

–

Development of lithography with 20nm features FAR in future!!!

• Novel Approach to IC-based rewritable mass storage

–

Disk drives define bit cells by mechanical positioning of head

–

IC-Based equivalent - newly emerging research suggests

MEMS-based positioning mechanisms

Probe-tip-based reading and writing mechanisms for data storage

(e.g., Scanning Tunneling Microscope - STM, AFM, MFM, etc.)

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 10 September 1999

&DUQHJLH 0HOORQ

Probe-Based Storage Mechanisms

• Rewritability and fast write time are critical

– Cache applications may write more often than they read

due to speculative prefetching

• Rewritable storage mechanisms

– Thin-film magnetic recording

» Continuous-film bit size > 20nm dia. marks

(thermal stability limit)

» Tips / write heads are quite difficult to design

» Discrete patterned media could result in smaller marks

– Amorphous / crystaline material (e.g., rewritable CD)

» Requires “thermal” generation

» Tips / write heads / read heads straightforward

» Mark size related to tip and thermal properties - order 3-10nm dia.

– Ferro-electric materials

» Electric field causes transformations

» bv

» Mark size related to tip shape - order 10-20nm dia.

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 11 September 1999

&DUQHJLH 0HOORQ

Probe-Based Storage Mechanisms

• Rewritability and fast write time are critical

- Approaches:

– Thin-film magnetic recording

» Continuous-film bit size > 50nm dia. marks

(by thermal stability & SNR)

» Tips / write heads are quite difficult to design

– Amorphous / crystaline material (e.g., rewritable CD)

» Tips / write heads / read heads straightforward - roadblock rewritability

» Mark size related to tip and thermal properties - order 3-10nm dia.

– Ferro-electric materials

» Electric field causes transformations - simple probe tips

» Mark size related to tip shape and media structure - order 10-20nm dia.

» Main roadblock: rewritability and media structure control

– Thermal melting of materials (IBM Almaden / Zurich approach)

» Melt plastic with hot probe tip

» Mark size related to tip shape - order 30-40nm dia.

» Main roadblock: Block erase to re-melt large area

– Charge storage on insulators (like EEPROM)

» Easy to fabricate - big concern is long term charge stability

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 12 September 1999

&DUQHJLH 0HOORQ

Magnetic Storage

Magnetic Probe tip

Magnetic media

•

Magnetic storage is an attractive choice

– > 1 GB / cm

density achievable today

– Reversible (rewritable) - unlimited # of cycles

– Proven to be robust

– Stable

– Most well-understood technology

•

Other possibilities should be explored too

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 13 September 1999

&DUQHJLH 0HOORQ

Magnetic Read-Write Mechanisms

• Reading mechanisms:

Soft Tip

Medium

Magnetic Flux

Tip Actuation

Write

Drive Coil

Lines

Current

Medium

Sense

Magnetic Tip

Current

Spin

Tunneling

(electrically
conductive)

Tip

Actuation

Dependent

Current

Medium

Magnetic

Sense

GMR

Current

Sensor

Tip

Actuation

Flux Lines

(soft magnetic tips)

magnetoresistive sensor

spin-dependent tunneling

(requires separate erase head)

• Writing mechanisms:

Hard Tip

Medium

Magnetic Flux

Tip Actuation

Z - motion at bit rate

Lines

hard magnetic tip

Z - motion flat

Medium

Sense Current

Magnetic Tip

Force

(electrically

conductive)

Tip

Actuation

Magnetic

MFM tip / current read

Hard Tip

Z - motion

at bit rate

Z - motion

flat

Z - motion

flat

Metallized

Resistive Medium

External

Tip

Write Current

Magnetic Field

Tunneling Current

Heated

Region

Tip

Actuation

thermo-magnetic writing

Magnetic Medium

Z - motion flat

inductive writing

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 14 September 1999

&DUQHJLH 0HOORQ

Example of Tip Fabrication

• Spindt tip process

• Pt Tips survive

post-CMOS
fluorine-based
plasma etches
(could be covered
with resist if
necessary)

• Multiple Pt tips

deposited on a
single CMOS
micromachined
structure

oxide

substrate

Tip Material

resist

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 15 September 1999

&DUQHJLH 0HOORQ

Positioning: The Rotating Disk

• Traditional Approach in Disk Drives

– Theta position provided by rotation

– Radius position controlled by macroscopic magnetic actuator

– Z position (height) controlled by air flow under air bearing surface

Velocity = ( 2

ππππ

R ) x RPM / 60

Rotational Latency = 60
2 x RPM

Data Rate = Lin. Density x Velocity

BAD NEWS !!!
- No good bearings
in MEMS Technology

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 16 September 1999

&DUQHJLH 0HOORQ

MEMS Actuator Design

• Consider 1D example

• Use “beam” springs

– Avoid friction and contact

– Repeatable and reliable

– K

∝

∝ ((((

f / L)

– Decreasing feature size (f) -- f

• Use electrostatic actuation

– Easily integrated with electronics

– Electrostatic force, F

∝

L V

• High-voltage IC process

– improves force as V

• Decreasing IC feature size (f)

– improves force as V

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 17 September 1999

&DUQHJLH 0HOORQ

MEMS Actuator Performance

• Three primary performance measures for actuator

– % swept length (area)

– Seek Time

– Max Operating Shock V = 10V f = 2

µµµµ

m 2

µµµµ

m 1

µµµµ

• For 1-D example:( f is feature size)

L= 1mm 0.5mm 0.25mm

• % swept length

≅ ≅

26 % 3.2% 12.5%

οοοο

2 E

• Access time

≅≅≅≅

≅

12ms 2.1ms 1.1ms

ρρρρ

• Max G Shock

≅

4.6 g 9.2g 37g

(10% force)

οοοο

ρρρρ

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 18 September 1999

&DUQHJLH 0HOORQ

Many Stage Possibilities

• What Moves??

– Move the media in X & Y, move the head in Z

» PRO:

Single media can have single actuator and suspension for

many probe tips -- very high storage density

» CON:

Large media sled is very massive -- long access times and

poor shock resistance

– Move the head in X & Y & Z

» PRO:

Head mass is very small -- fast access times and

excellent shock resistance, many simultaneous accesses possible

» CON:

Swept area limited by single actuator overhead

(probably not better than 1%) -- low storage density

– Compromise: Move the M small media sleds in X & Y, move head in Z

» PRO:

Smooth trade-off between swept area and access time / shock

resistance

» CON:

More complex control systems, fewer simultaneous accesses

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 19 September 1999

&DUQHJLH 0HOORQ

MEMS + Probe-Tip-Based Storage

•

X-Y Translation

– Avoids rotation problem
– No bearings

•

Typical MEMS electrostatic
actuators inefficient

– Many different structures
– Best efficiency 10%

•

Conclusion:

–

Array of Probe Heads

–

Media long
compared to
actuators

•

Large Array of
Probe Heads:

– Media moves by probe tip

pitch in X and Y

Tip

Array

Media

Coating on

Bottom

Surface

1 cm

1 mm

Electronics

Media

Sled

Media

Sled

∆∆∆∆

Y < Yo / 10

Actuator

Size

Media

Motion

Example Device

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 20 September 1999

&DUQHJLH 0HOORQ

Media Actuator Examples

•

Kionix (Cornell) stage - 4mm x 4mm with 40 µm stroke

•

HP Design Example: 1mm x 1mm media, ±8 µm stroke
2mm x 2mm total size

⇒

25% swept area

–

Hoen et. al., “Electrostatic Surface Drives, …,” Transducers 1997

•

CMU Design - 8 mm x 8 mm media, ±50 µm stroke
14mm x 14mm total size

⇒

32.7% swept area

electrostatic
x actuator

electrostatic
y actuator

beam spring

CMOS chip
standoff

and sensor

media

A’

CMOS die

moving stage

tip-media
spacer

anchor point
of media stage

handle
wafer

8 mm

oxide
insulating
layer

IC package

tips

media

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 22 September 1999

&DUQHJLH 0HOORQ

Integration of Circuits with Tips

•

Typical system has large # tips; e.g., 6,400. Each must be:

–

Sensed and servoed in Z and

∆∆∆∆

–

Data must be read and written

–

Too many wires unless some electronics integrated with probe tips

y servo

x servo

tip write

tip read

media

tip

controller

track
error

data

address

z servo

tip servo

mux

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 23 September 1999

&DUQHJLH 0HOORQ

CMOS High-Aspect-Ratio MEMS

Overglass

3-metal
interconnect

0.5 µm
n-well
CMOS

Gate polysilicon

Silicon
substrate

■

Structures made using conventional CMOS

■

Hewlett-Packard CMOS14 available from MOSIS

■

Construct a beam from CMOS chip:

* G.K. Fedder, et.al., Sensors and Actuators, vol. A57, (1996).

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 25 September 1999

&DUQHJLH 0HOORQ

CMOS High-Aspect-Ratio MEMS

" 1.2 µm

~ 5 µm

Composite
beam

Etched pit

Stator
electrodes

Metal
layers

■

Need 1000’s of wires from MEMS/probes to electronics

■

Developed CMOS compatible MEMS process at CMU

■

Post-CMOS etch of silicon substrate

■

Structures are undercut & released

* G.K. Fedder, et.al., Sensors and Actuators, vol. A57, (1996).

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 26 September 1999

&DUQHJLH 0HOORQ

CMOS High-Aspect-Ratio MEMS

" 1.2 µm

~ 5 µm

Composite
beam

Etched pit

Stator
electrodes

Metal
layers

■

Need 1000’s of wires from MEMS/probes to electronics

■

Developed CMOS compatible MEMS process at CMU

■

Post-CMOS etch of silicon substrate

■

Structures are undercut & released

* G.K. Fedder, et.al., Sensors and Actuators, vol. A57, (1996).

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 27 September 1999

&DUQHJLH 0HOORQ

FETs and Circuits on MEMS Chip

• Isotropic silicon

etch required to
release MEMS
structures

• FETs closer than

8 µm are etched
away

• Beyond 8 µm,

FETs’ gain and
threshold are not
affected

• No degradation

after 4 months

7um

. . . . . .

12 test FETs

actual edge of etched pit

metal opening for etching

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 28 September 1999

&DUQHJLH 0HOORQ

Example of Z Probe Tip Actuator

• Z suspension needed

for every Probe Tip

– surface roughness
– surface curvature
– tip height variations

• IBM Zurich - AFM tip array

– 32 x 32 array of probe tips
– Despont et. al., MEMS ‘99

• Alternative: active Z actuation

parallel-plate metal-to-media

– Allows active spacing
– Does not require contact

• Small Lateral Deflection

– Thermal expansion mismatch

• CMU Design Example

spring
suspension

structure length = 231 µm,

beam width = 1.8 µm,

comb gap = 1.8 µm

lateral
comb actuator

vertical
actuator

probe tip
pad

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 29 September 1999

&DUQHJLH 0HOORQ

New Degrees of Freedom

• Hopefully, convinced IC-based mass storage possible

– Will require focussed research to make it practical

• IC-Based storage has many degrees of freedom; e.g.,

– What is storage mechanism - magnetic is best candidate so far?

– How are reading and writing performed?

– How many probe tips on a single MEMS Z-suspension?

– How many probe tips per MEMS media sled?

– How many instances of detection and positioning electronics?

– How many media sleds / IC?

– Is processing integrated with storage subsystems?

If so, what is the best pattern for interconnection?

– What kind of processors should be integrated with storage subsystems?

Traditional DSPs vs. VLIWs vs. FPGAs vs. something new?

• Understand applications to make informed trade-offs

– We need to study and understand broad application classes

– We need simulation model to relate application needs to design choices

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 30 September 1999

&DUQHJLH 0HOORQ

Parallelism of Tips for Data Rate

• Example of Decision -- pattern for reading and writing

– More tips active more data rate and more power dissipation

– Cannot access independent locations with tips on same media sled

Read from all

tips in the

row in parallel.

80 bits per

read cycle

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 31 September 1999

&DUQHJLH 0HOORQ

Data Assignment Decisions

• Another degree of freedom is how to map data to device

The media sled moves in

X and Y above the tip array.

Tips are fixed so they each

are positioned over the same

track group in each square.

However, each tip has some

freedom to move in the X direction

so they can

deflect +/- 5 tracks within

a track group, independent of the other tips.

Track Group

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 32 September 1999

&DUQHJLH 0HOORQ

Example Data Access Pattern

• Even choices to make for how to scan sled across data

– And, these choices all make a difference to application performance

Tips in row 1 read down

to end of track 2.

Tip switch - row 2 tips

turn on.

Tips in row 2 read up to

end of track 2.

Sled seeks to next

track.

Tips in row 2 read down

to end of track 3.

Tip switch - row 1 tips

turn on.

Tips in row 1 read up to

end of track 3.

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 33 September 1999

&DUQHJLH 0HOORQ

Exploration Via System Simulation

• Created Processor+Storage System Model using SimOS

– Work of Professors Nagle and Ganger and their students

– Implemented a simple MEMS model based on original SimOS disk model.

– Implemented a new SimOS disk model using DiskSim.

– Re-implement a detailed MEMS model as a module in DiskSim.

• Resimulate System Performance

with different MEMS Storage System parameters

– Can see application-level impact of changes in basic science

– For example,

» Physicist working on new head/media combination with

higher density but longer write time can see impact on system

• This overall flow of information is at core of center idea

– Basic Science

→

Storage subsystem designer

→

Technology

– Technology + architectural decisions

→

proposed new system

– Proposed new system + simulation

→

feedback to all levels

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 34 September 1999

&DUQHJLH 0HOORQ

Example Simulation Results

•

Capacity

–

2.56 GB per cm

with 1 media actuator with 1cm x 1cm sled

(about the worst possible seek time case)

•

Latency

–

Tip actuator takes 5 ms to 0.1 ms to seek

Compare to 12 ms to 0.6 ms for Seagate Cheetah Disk

–

MEMS tip actuator can read from ±5 tracks without media seek

–

MEMS Media sled inertia dominated - 1 to 10 ms extra latency
exact motion can be predicated given a maximum actuator force

•

Data Rate

–

Assume Probe Tip recording at 200 kb/s per tip

–

100 to 10,000 simultaneous tips depending on power dissipation limits

–

2Gb/s maximum possible for 1 cm

device

•

Power

–

Dominated by number of active tips

–

Assume 5 mW / active tip + 100mW for media positioner

L. R. Carley - “Highly Integrated Information Processing and Storage Systems” page 37 September 1999

&DUQHJLH 0HOORQ

IC-Compatible CPU +Mass Storage

• IC-Based Data Storage is Achievable

– Lithographically defined approaches already commercially viable

» EEPROM, FeRAM, MRAM (coming soon)
» Storage capacity limited by photolithographic resolution

– X-Y-Z actuation with CMOS IC compatible MEMS

» Demonstrated small swept area
» Much research needed to increase swept area

– Demonstrated simple STM tip fabrication

» Need to develop process compatible magnetic “head” designs

– Demonstrated simple carbon film WORM media

» Need to develop process compatible magnetic media designs

– System Demonstrations

» Need to integrate components with electronics into system demos
» Currently working on a basic Z-only MEMS magnetic probe demo

• Why Do It ??

– Remember Advantages of IC-Based Computing + Storage Systems

» Low Cost, Short Access Time, Low Volume and Mass,
» Potential for Archivability of Data / Software / Hardware