UIC_Thesis_Defense_Vinella
-
Upload
paolo-vinella -
Category
Documents
-
view
21 -
download
1
Transcript of UIC_Thesis_Defense_Vinella
A study of using STT-MRAM as Memory PUF: design, modeling and quality evaluation
Paolo Vinella
– Thesis Defense –
Academic Committee:Wenjing RaoZhichun ZhuFabrizio Bonani
Chicago, Turin - May 7, 2015
Master of Science in Electrical and Computer Engineering
SummaryCONTENTS:
• Introduction on the relevant topics
• Thesis goals & Workflow organization
2
Introduction on the relevant topics
3
• Preliminary work on MRAM to realize “binary signature”
• Low-cost, secure, reliable and tampering-resistant
• Applications: chip signature + secure authentication
New memory device: MAGNETIC MEMORY
(STT-MRAM)
New application for Hardware Security: PHYSICALLY UNCLONABLE FUNCTION
(PUF)
Summary
Thesis goals & Workflow organization
4
STT-MRAM tech PUF
Tunneling Magnetoresistance
Previous Work: MRAM as PUF
MY
WO
RK 2 – Verify and re-implement previous work+
Augment previous work
4 – Propose a new “Strong PUF”
1 – Study process variation effect on MRAM cell
3 – Evaluate PUF Quality
Summary
MRAM Technology
CONTENTS:
• ITRS: need of new memory tech
• MRAM a good new memory candidate
• STT-MRAM working principles
• Magnetoresistance effect
• MRAM Memory Array
• Applications
5
ITRS: need of new memory tech
6MRAM Technology
• International Technology Roadmap for Semiconductors, 2013 Issue
• Need to solve ICs scale-down limits & limitations
• NEAR TERM (2013-2020):o CMOS scalingo Enduranceo Noise margin reductiono Solve memory latency gap
• LONG TERM (2021-2028):o New memory structures: replace DRAM/SRAMo Non-volatile memories (NVM) possible candidateo Implement non-charge-storage type of NVM cost effectivelyo Keep it smart: “low-cost, high-density, low-power, low-latency”
MRAM a new good memory candidate
7MRAM Technology
• Current memory technologies: none is the best!• Advanced MRAM (p-STT-MRAM) a solution
DREAM-MEMORY:ü unique, universal, fast memoryü only advantages of current memoriesü little cost / extra design effort
DREAM MEMORY
LATENCY
CAPACITY
1ns
10ns
100ns
1 ms
MB GB TB
Hard Disk drives
Flash (NOR)
Flash(NAND)
T R E N D
Field MRAM
CPU Registers
Cache (SRAM)
Main mem. (DRAM)
STT-MRAM
STT-MRAM working principles
8MRAM Technology
• Magnetic memory: based on ELECTROMAGNETISM
• M (magnetic dipole moment): natural magnetization of a material
• Based on MTJ: Magnetic Tunneling Junction
(Ferromagnet + Oxide + Ferromagnet)
FREE layerMgO
PINNED layer
• MRAM use = control M direction of ferromagnetic layer (“SWITCHING”)
o Field MRAM: use external magnetic field
o STT-MRAM: more clever, uses current flowing through
Magnetoresistance effect
9MRAM Technology
FerroMagnet
Tunneling Oxide
FerroMagnet
Rp
Rap
bit ‘0’
bit ‘1’
PARALLEL STATE (P):
ANTIPARALLEL STATE (AP):
MRAM Memory Array
10MRAM Technology
WL
BL
SL
Ip Iap
RMTJFREE layer
MgO
PINNED layer
RMTJ
R P
R APVMTJ
VMTJ
Source Drain
Gate
MTJ
G
D
S
(a) (b) (c)
BL j SL j
WL i
WL i+1
BL j-1 SL j-1 BL j+1 SL j+1
10 1WL i
BL j BL j+1BL j-1
APP APPROGRAMMING PATTERN:
+ V cell j-1 – – V cell j + – V cell j+1 +
I P I AP I AP
Applications
11MRAM Technology
NANO-ELECTRONICS• MTJ-based BJT• Pseudo-spin-MOSFET *
AS PROPER MEMORY
• Universal memory: p-STT-MRAM by Toshiba *
DIGITAL BLOCKS
• Unbalanced MTJ flip-flop à FPGA• Logic in Memory (runtime reconfig. LUT) *• Associative Computing
MICROARCHITECTURES• Normally-Off Processor with p-STT-MRAM (cache)• Resistive Computation: entire non-volatile CPU
à Hardware Security and PUF: the newest and less mature use ß
HW Security& PUF
CONTENTS:
• Chips Piracy
• Hardware Security
• Physically Unclonable Function (PUF)
12
Chips Piracy
13HW Security & PUF
• Engineering products = result of clever insights and breakthroughs
• Every idea should be protected against theft and plagiarism
• Fab-less company à chip design shared with manufacturer(s)
• Untrusted parties may access, steal, re-use design illegally
Company A(Fab-less Designer)
Company B(Manufacturer)
Unprotected Chip Design
Company A(Fab-less Designer)
Company B(Manufacturer)
Protected Chip Design
Hardware Security
14HW Security & PUF
• Fight design thefts: add some sort of extra “layer of security”
• Rules + protocols + implementations embedded at design stage of a chip
• New research area: HARDWARE SECURITY
• Identify a given chip: “ID, UNIQUE SIGNATURE”
Company A(Fab-less Designer)
Company B(Manufacturer)
Unprotected Chip Design
Company A(Fab-less Designer)
Company B(Manufacturer)
Protected Chip Design
Physically Unclonable Function (PUF) (1)
15HW Security & PUF
NEW BUT… OLD!
Identification of nuclear weapons during the Cold War:
1. “CHALLENGE”: spray thin coating light-reflecting particles on nuclear weapon
2. Particles are randomly distributed
3. “RESPONSE”: illuminate weapons from different
angles à get different interference patterns
UNIQUE/UNCLONABLE SIGNATURE STORED IN A SECURE DATABASE!
Physically Unclonable Function (PUF) (2)
16HW Security & PUF
• Physical: rely on a feature of materials constituting an electronic chip
• Unclonable: unique and tampering-resistant
• Function: generate useful information for identification purposes
• Exploit manufacturing process inaccuracies (rather fought at the digital level!)
• Challenge: let analog and digital domains coexist + using process variation
PROCESS VARIATION(oxide thickness, doping fluctuations,
feature size, trapped charges,…)
MANIFACTURING PROCESS STEPS
Chip#1
CHIPS LOTVariation in physical features of each chip
PUFyi = f x i
i=1…N
SIGNATURES
Easy to generate !
Difficult to invert "
Signature#1
Signature#2
Signature#i
Signature#N
…
…
x1
x2
xi
x N
y 1y 2
yi
yN
Chip#2
Chip#i
Chip#N
OXIDE
DOPING
FEAT. SIZE
TRAPPED
CHARGES
Physically Unclonable Function (PUF) (3)
17HW Security & PUF
• Totally random, easy to generate, hard to invert
• Analog PUF: laser on 3D token à unique interference pattern
• Popular digital PUFs: “Ring Oscillator ” and “SRAM initialization”
• Current practice for HW authentication: secret key in NVM
(EEPROM), and use hardware cryptographic operations
• PUF avoids to use EEPROM / other expensive HW: nothing
stored in a memory, but derives a signature “on the fly”
Previous WorkMRAM as PUF
CONTENTS:
• MRAM Resistance dispersion as PUF
• Previous Work ideas
• Sensing schemes
• Results
18
Ref: Zhang L. et al. , Feasibility study of emerging non-volatile memory based physical unclonable functions. In Memory Workshop (IMW) 2014 IEEE
MRAM Resistance dispersion as PUF
19Previous WorkMRAM as PUF
R [Ω]μ Rp
σ Rp
σ Rap
R [Ω]Rpnom Rapnom μ Rap
bit ‘0’ bit ‘1’bit ‘0’ bit ‘1’
IDEAL REALPROCESS'VARIATION
n"# $ Rp
Engineer resistance dispersion:
random physical phenomena à source of a robust signature generation
Previous Work ideas
20Previous WorkMRAM as PUF
For each WORD, process each BIT exploiting MTJ resistance variation
generating a PUF response bit
Response is a binary string
STT-MRAM
chip
WLi
BLj SLjRMTJ
B I T L I N E S ( B L + S L )
W O
R D
L
I N
E S
b0 b1 … … … … bL
… … … …
PUFyi = f x i
i=1…L
TO THE INSIDE …
PUF Signature String 0PUF Signature String 1
…………
PUF Signature String M
B I T L I N E S ( B L + S L )
W O
R D
L
I N
E S
L - bits
M -
word
s
Sensing schemes
21Previous WorkMRAM as PUF
+
–
Ref_CellData_Cell
ΔIsens,min
IDATA
IREF+
–
Data_Cell_2Data_Cell_1
ΔIsens,min
IDATA1
IDATA2
WLi WLi
WLi WLi
PUF_bit PUF_bit
CASE 1 CASE 2
• Ideally, two identical currents (nominal RP or RAP)
• Process variation à resistance dispersion leads to different actual values
• Same read voltage à to two different currents!
• Sense amplifier detects mismatch and produces output PUF bit
Results
22Previous WorkMRAM as PUF
AS A PUF
• STT-MRAM can generate a “good quality” PUF
• Best results with differential schemes (no reference)
• STT-MRAM the best NVM as PUF (BER)
AS A MEMORY
• “Dream Memory”
• MTJs have high endurance à higher signature reliability
• Resistant to external disturbances
23
MY WORK1. Study on process variation effects on
MRAM cell
2. Verify + Augment previous work
3. Evaluate PUF Quality
4. Outlook: upgrading to a “strong PUF”
The complete simulation framework, overview
24
MAIN_RpRap_analysis.m
MTJ_Rp_Rap_populations.m
tox_var_effect_on_RpRap.m
MAIN_PUF_simulation.m
case1_1T1MTJ_with_reference.m
case2_2T2MTJ_no_reference.m
case3_2T2MTJ_with_reference.m
Case4_2T2MTJ_no_reference.m
BER_for_various_deltaI_PvsAP.m
case1_UniquenessEval.m
case2_UniquenessEval.m
case3_UniquenessEval.m
case4_UniquenessEval.m
sense_amplifier.m
plot_BER.m
1N=5000tox_nom=0.85nm, σ=1%RA=10Ωμm2
A=B=65nm, σ=3%TMR0=120%Vbias_nom=0.1V, σ=1%
AB
tox0.1V
N, tox_nom, tox_dev,RA, A, B, size_dev, TMR0,Vbias_nom, Vbias_dev
N, Rp_pop, Rap_pop,Vbias_pop, tox_nom, RA, A, B, TMR0
N, RA, A, B, TMR0, Vbias_nom
Rp_pop, Rp_pop,Vbias_pop
FIVE PLOTS ONPOPULATION PERFORMANCES
@ tox = 0.85 ; 1 ; 1.15 nmPLOT : σRp , σRap = f ( σtox )
μRp , μRap = f ( σtox )Given : Vbias, A, B constant and varying
MTJ_pop_param_eval.m
used by
used by
used by
2choose: !
σtox=1%σFeatSize=3%
then:Nchips=100PUF size: (m=32) x (l=32)
other param. exactly as
CHIPS_STT_MRAM
L
M
Nchips
1
error,Iplus, Iminus resp_bit
used by
used by
used by
used by
used by
N, m, l, state,CHIPS_STT_MRAM,sense_ampl_err
same as case1
same as case1
same as case1
Menu, N, m, l, CHIPS_STT_MRAM,tox_nom, A, B, TMR0
flag=1
flag=1
flag=1
flag=1
flag=0
IF flag=1 && sense_ampl_err=0
IF flag=1 && sense_ampl_err=0
IF flag=1 && sense_ampl_err=0
IF flag=1 && sense_ampl_err=0
err_bit_percentage
err_bit_percentage
err_bit_percentage
err_bit_percentage
FOR THE SELECTED CASE, let sense ampl. error varythen compute and plot the associated BER.
PLOT: Hamming Distance, Hamming Weight, Bit Aliasing
PLOT: Hamming Distance, Hamming Weight, Bit Aliasing
PLOT: Hamming Distance, Hamming Weight, Bit Aliasing
PLOT: Hamming Distance, Hamming Weight, Bit Aliasing
PLOT the BER for Case 1, 2, 3 and 4
o Implemented in Matlabo Made versatile
Study on process variation
effects on MRAM cell
CONTENTS:
• Memory cells MTJ resistance model
• RP/RAP interesting results
• Oxide thickness effects on RP/RAP
populations
25
Memory cells MTJ resistance model
26Process variation effects on MRAM cell
• MRAM as PUF à RP and RAP
• Nominal values with equation models, most accurate found (*):
R#,R%# = f (t+, ,RA , Area , ϕ , V3456)
(*) Ref: Zhao W. et al. , Macro-model of spin-transfer torque based magnetic tunnel junction device for hybrid magnetic-CMOS design. In Behavioral Modeling and Simulation Workshop 2006, IEEE
• So far, dispersion of RP / RAP found producing sample devices
• Bond the two approaches to check suitability as a PUF
• Simulation based on Monte Carlo to emulate process variation effects on RP/AP
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.52000
2500
3000
3500
4000
4500
5000
5500
6000
6500
7000
Rp and Rap vs Bias Voltage of MTJ, with no variations in device geometrytox = 0.85 nm − Feat.size = 65 nm − TMR = 120 %
Vbias [mV]
Rp
and
Rap
[Ω]
Rp=f(VMTJ)Rap=f(VMTJ)
RP/RAP interesting results
27Process variation effects on MRAM cell
2000 3000 4000 5000 6000 7000 8000 90000
50
100
150
200
250
300
350
400
450
500
Rp and Rap MTJ simulated distribution − Sample size = 5000tox = 0.85 nm − Feat.size = 65 nm − TMR = 120 %
Resistance [Ω]
Co
un
ts [
# d
evi
ces
with
th
at
Re
sist
an
ce]
Rp distributionRap distribution
INPU
TSN=5000 cells, elliptic shape
PARAMETER μ σtOX 0.85 nm 1 %
A , B 65 nm 3 %Vbias 100 mV 1 %TMR0 120 %
RA 10 Ωμm2
Oxide thickness effects on RP/RAP populations
28Process variation effects on MRAM cell
• Understand how far to push resistance dispersion of MTJs
• PROCESS VARIATION à the base for for PUF
• At same time, need reliable memory cells for classic applications: data storage!
Oxide thickness the center of gravity
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
5
10
15
20
25
30
35
40
45
50
STT−MTJs σRap
and σRap
, (% over mean value) − Sample size = 5000
TMR = 120%V
read = 0.1V , Feat.size = 65nm (both w/ 5% variation)
For different oxide thicknesses
σtox
[%]
σR
p ,
σR
ap
[%]
σRp
@ tox
= 0.85nm
σRap
@ tox
= 0.85nm
σRp
@ tox
= 1nm
σRap
@ tox
= 1nm
σRp
@ tox
= 1.15nm
σRap
@ tox
= 1.15nm
• tOX ↑ and σtox ↑à RP/RAP dispersion ↑
• No variability whatsoever à back to
nominal case
• But when tOX<1nm also variability of other
physical quantities matters
• Strong influence of tOX on RP/RAP: makes
sense (e- tunneling)
Verify + Augment previous work
CONTENTS:
• Augment MRAM PUF Signature
• Design schemes variation
• Simulation details
29
Augment MRAM PUF Signature
30Verify+Augment previous work
STT-MRAM
STT-MRAM
STT-MRAM
STT-MRAM
STT-MRAM
B I T L I N E S ( B L + S L )L - bits
PUF Signature for chip i
(entire memory matrix used) W
O R
D
L I N
E S
M -w
ords
MRAM_CHIP_N
MRAM_CHIP_1
MRAM_CHIP_2
MRAM_CHIP_i
MRAM_CHIP_N-1P
R O D
U C
T I O
N
L O
T
STT-MRAM
STT-MRAM
STT-MRAM
STT-MRAM
STT-MRAM
STT-MRAM
N devices in theproduction lot
N chips of STT-MRAM used for a par ticular memory-
related purposeM
A N
I F
A C
T.
Design schemes variation
31Verify+Augment previous work
WLi WLi
CASE 1 CASE 2
+–
ΔIsens,min
PUF_bit
AP or PAP or P
+ –
ΔIsens,min
PUF_bit
AP or PAP or P
IDATA IREF IDATA1 IDATA2
\WLi WLi
+ –
ΔIsens,min
PUF_bit
CASE 3
AP P
Δ I 1
AP P
Δ I 2
CASE 4
IDATA1 IDATA2 IREF2IREF1
+ –
ΔIsens,min
PUF_bit
AP P
Δ I 1
AP P
Δ I 2
IDATA1 IDATA2 IDATA4IDATA3
• Trend for MRAM to store data: differential approach
• Memory size (number of MTJs) increases
Simulation details
32Verify+Augment previous work
• Generate a PUF Signature of (32x32) bits for each chip
• Production lot: N=100 different chips
• Simulator steps:
1. Generate population of STT-MRAM chips à a 3D matrix
2. Generate PUF response string for each Case 1-2-3-4
3. Evaluate PUF Quality: “Uniqueness” and “Reliability”
L
M 0 1 1 0 0 0 0 0 0 1 0 … … … … 1 0 0 0 1 1 1
M x L = Z
PUF binary response string R i=r 1∥r 2∥...∥r j∥...∥r Z in vectored form:
0 0 1 1 0 1 10 0 1 0 0 1 10 0 1 0 0 1 01 1 1 0 1 1 11 1 0 0 1 0 10 1 1 1 1 0 11 0 0 0 1 1 1
1 1 0 1 1 0 10 0 1 0 0 1 10 0 1 0 0 1 11 1 1 0 1 1 11 1 0 0 1 0 10 1 1 1 1 0 01 0 0 0 1 1 1
1 1 1 1 0 1 00 0 1 0 0 1 10 0 1 0 0 1 01 1 1 0 1 1 01 1 0 0 1 0 00 1 1 1 1 0 01 0 0 0 1 1 0
0 0 1 0 1 0 10 0 1 0 0 1 00 0 1 0 0 1 11 1 1 0 1 1 11 1 0 0 1 0 00 1 1 1 1 0 01 0 0 0 1 1 1
1 0 0 0 1 0 10 0 1 0 0 1 00 0 1 0 0 1 01 1 1 0 1 1 01 1 0 0 1 0 10 1 1 1 1 0 11 0 0 0 1 1 0
0 1 1 0 0 0 00 0 1 0 0 1 00 0 1 0 0 1 11 1 1 0 1 1 11 1 0 0 1 0 00 1 1 1 1 0 01 0 0 0 1 1 1
RESHAPE AS A VECTOR
Ni=1÷N
* mem. matrix M x (4 x L) x N
Evaluate PUF Quality
CONTENTS:
• PUF Signature Quality metrics
• Uniqueness Results
• BER (Reliability) Results
• Influence of reading voltage over BER
• Keep increasing reading voltage then?
33
PUF Signature Quality metrics
34Evaluate PUF Quality
• UNIQUENESS: expresses randomness of the signature, how unique it is w.r.t. others
o Hamming Distance
o Hamming Weight
o Bit Aliasing
1 1 0 0 1 0 0 1 … … … … 0 0 1 N0 0 1 1 1 1 1 0 … … … … 1 1 11 1 0 0 0 1 0 1 … … … … 1 0 00 1 1 0 0 0 1 0 … … … … 1 0 1
=?
0 1 1 0 0 0 1 0 … … … … 1 0 1
1 1 0 0 1 0 0 1 … … … … 0 0 1 N0 0 1 1 1 1 1 0 … … … … 1 1 11 1 0 0 0 1 0 1 … … … … 1 0 00 1 1 0 0 0 1 0 … … … … 1 0 1
• RELIABILITY: stability over undesired parameter fluctuations (voltage, temperature)
o BER: Bit-Error Rate, due to sense amplifier quality and reading voltage
Uniqueness Results
35Evaluate PUF Quality
• All cases provide nominal HD/HW/BA close to 50% (ideal value)
• Avoiding the use of a reference cell improves Uniqueness (better randomness)
• Although Case 3 (and especially 4) use far more cells, Uniqueness not penalized
BER (Reliability) Results
36Evaluate PUF Quality
• P state better! Lower BER (lower R à higher currents à amplifier more relaxed)
• Case 3 & 4 are a “reversed-combo” (P+AP) à BER still better than 1 & 2
• Also, we showed that BER ∝ ΔISENS,MIN
CASE 1 CASE 2 CASE 3 CASE 4 CASE 1 CASE 2 CASE 3 CASE 4
Influence of reading voltage over BER
37Evaluate PUF Quality
• PUF stability changes with voltage / temperature fluctuations (high “variations”!)
• Model used does not provide temperature dependence à focus on voltage only
• VBIAS affects RP and RAP à effects on BER
2000 3000 4000 5000 6000 7000 8000 9000 100000
0.5
1
1.5
2
2.5
3
3.5x 10
4
Rp and Rap MTJ simulated distribution − Sample size = 409600
@ READING VOLTAGES: Vbias = 50mV ; 100mV ; 150mV
tox = 0.85 nm − Feat.size = 65 nm − TMR = 120 %
Resistance [Ω]
Counts
[# d
evi
ces
with
that R
esi
stance
]
RpRpRpRapRapRap
BETTER
BETT
ER
Keep increasing reading voltage then?
38Evaluate PUF Quality
• Higher reading voltage, lower BER but… there is a limit!
o Power dissipation, especially if PUF response generated multiple times @runtime
o AP and P distributions will start to overlap: BER also as a memory!
o Memory cell MTJs could switch magnetization: unwanted write operation
Outlook: upgrading to
a “Strong PUF”
CONTENTS:• Motivations to enhance a PUF• Weak PUF vs Strong PUF• Re-engineering MRAM as Strong PUF• Memory layout assumptions• A proposed implementation of Strong
MRAM PUF• Complexity for the attacker Φ
39
Motivations to enhance a PUF
40Propose new “Strong PUF”
• Unique signature à single binary stream, can be used as cryptographic processes
• Example: secret key to encrypt/decrypt data on a device
• This information must thus be kept private
• Sometimes we need authentication
• Data exchanged with a device to be identified as valid
• Need of a key that might travel through unsecure media
Weak PUF vs Strong PUF
41Propose new “Strong PUF”
• Weak PUF: a limited set of challenge-response pairs (CRPs)
o So far, our MRAM PUF in this category. Challenge: WL à Response: binary string
o Application: unique Signature
• Strong PUF: a high number of CRPs
o Back to the origins (cold war)
o Application: authentication (secure server +chip at potential unsecure location)
Re-engineering MRAM as Strong PUF
42Propose new “Strong PUF”
• Previously studied Case 3 & 4 based on cells grouping
• Idea can be augmented: digital switches connecting MRAM cells to sense amplifier
• Now challenge is WL + a variable “command” driving switches
B I T L I N E S ( B L + S L )L - bits
(ch-resp)1
(ch-resp)2(ch-resp)3
…(ch-resp)C
W O
R D
L
I N E
SM
-wor
ds
FOR EACH ROW: use switches to realize connections, grouping different
memory cells with sense amplifiers:
PUF bit
⇐
Original STT-MRAMmemory matrixin a given chip
Interconnections overhead
Sense Ampl. row
A large enough number of possible (challenge-
response) pairs
Designer chooses differentswitches patterns leading to several (challenge-response) pairs
Memory layout assumptions
43Propose new “Strong PUF”
B I T L I N E S ( B L + S L )L - bits
W O
R D
L
I N E
SM
-wor
dsELEMENTARY MEMORY CELL
Sense Ampl. row
MEMORY MATRIXBL j SL j
WL i = 1 I cell
V DD
L/2 SENSE AMPLIFIERSserving couples of cells for the i-th rowselected activating its respective WL-i
+ –
ΔIsens,min
A proposed implementation of Strong MRAM PUF
44Propose new “Strong PUF”
• Decoder bit stream can be stored in unused rows of MRAM à memory + PUF!
L bits
WL i⇒for each memory cell
+ –
log2(L)-to-L DECODER
log2(L)-bits stream
A B each group contains K elements
A DECODER for each memory cell, connecting it to any of the available sense amplifiers’ inputs
S available sense amplifiers (typically S=L/2)
PUF$Response$Binary$String
the word line activates a specific word, thenthe generation of the PUF response starts
The CHALLENGE is now the combination of the word line WLiand of the switches
programming pattern bits stream
The binary RESPONSE string
for the word line “i”
BL j SL j
WL i = 1 I CELL
VDD
+ –
C D
+ –
E F
+ –
… …
Complexity for the attacker Φ
45Propose new “Strong PUF”
• Resistance to Brute force attacks: how complex trying all combinations of challenges
• Strong PUF à complexity for attacker exponentially increasing with memory size
• Not a closed form à proved to beexponential evaluating up/low bounds
• L: # elementary memory cells per row
• K: # cells grouped to each sense amplifier input
A SEMILOG-Y PLOT!0 5 10 15 20 25 30 35
100
102
104
106
108
1010
1012
1014
1016
1018
1020
Memory matrix word width L
Logarith
mic
plo
ts o
f Φ
, M
IN a
nd M
AX
Semilog plot of attacker complexity Φ as function of memory word width L. Fix M=32 words
MIN
Phi
MAX
Conclusions,Future Work& Remarks
CONTENTS:
• Conclusions
• Future Work
• Acknowledgements
46
Conclusions
47Conclusions, Future Work & Remarks
• Very new memory applied to a new paradigm of Hardware Security
• STT-MRAM both good memory and good PUF
• Simulation framework can be further expanded
CHALLENGES OVERCOME
o Understand previous work, re-implement from scratch
o MRAM very new: unified / comprehensive literature not present yet
o PUF: nanoelectronics + analog + digital domains
o MRAM DC electrical models incomplete, never applied with process variation
MAIN CONTRIBUTIONS
1. Start from scratch to reproduce previous work
2. Twist of ideas, new models
3. Larger departure from original work (land on “Strong PUF”)
Future Work
48Conclusions, Future Work & Remarks
MRAM
• Overcome IC scale-down issues, speed, non volatile, low power
• Propose/agree on standards and design schemes
• Translate early prototypes in cost-effective innovative logic
• More accurate / comprehensive models and equations
PUF
• Expand Strong PUF (HW overhead vs speed,…)
• Implement real-time complete authentication system based on MRAM as PUF
• Use Fuzzy Extractor to generate a more stable PUF Response
Acknowledgements
49Conclusions, Future Work & Remarks
• Grateful for this research experience, both as RA and TA at UIC
• Staff and Faculty, from both UIC and PoliTo
• Advisors:
o Prof. Wenjing Rao
o Prof. Fabrizio Bonani
• Committee Members:
o Prof. Zhichun Zhu
• Soroush Khaleghi
Thank you!
Appendix
MRAM vs conventional memories - BENCHMARKS
52
Source: Latest Advances and Future Prospects of STT-RAM, Alexander Driskill-Smith, © 2010 Grandis Corporation
STT-MRAM vs Field MRAM
53
Source: Latest Advances and Future Prospects of STT-RAM, Alexander Driskill-Smith, © 2010 Grandis Corporation
MRAM vs conventional memories - BENCHMARKS
54
• ANISOTROPIC MAGNETORESISTANCE (1856): R = f (θ J,M)
(a) External+field+applied+→+PARALLEL++→+RLOW
(b) No+field+applied++→++ANTIPARALLEL++→+RHI
• (GIANT) TUNNELING MAGNETORESISTANCE (TMR Ratio) (1975):+FM//OXIDE//FM+(“MTJ”)
• GIANT MAGNETORESISTANCE - GMR (1988,+Nobel+in+2007):+FM//NM//FM+layer
FMNMFM
FMNMFM
Nowadays,+near+
TMR=500% at+room+
temperature!
“Pinned layer”+(fixed+direction)+++Oxide +++“Free layer”+(able+to+switch+magnetization)
Use+THIN films+for+abrupt+P/AP+
transition
Spin-Torque Transfer effect
55
• Giant Tunneling Magnetoresistance (1975): MTJ stack exhibits large associated resistance
• Combined with SPINTRONIC (1996): “spin-torque transfer ” effect
• Ferromagnet acts as a “polarization filter ” (alteringelectrons populations)
M1 fixed (pinned): minority-spin electrons
reflected back!
Scheme from Magnetic Memory: Fundamentals and Technology, Cambridge Press.
MTJ resistance model used
56
• Nominal values RP / RAP with equation models
• Most accurate found to be:
• Equations applied to derive the properties of the memory matrix cells (associated
resistances) after letting their physical inputs vary simulating the process variation.
Ref: Zhao W. et al. , Macro-model of spin-transfer torque based magnetic tunnel junction device for hybrid magnetic-CMOS design. In Behavioral Modeling and Simulation Workshop 2006, IEEE
Currents distribution: evaluating the PUF
57
• Good dispersion RP and RAP à better PUF uniqueness
• Normal distribution required (easy to study)
• RP has higher currents à hopefully lower BER
• Differential currents as high as possible to be correctly sensed
−2 −1 0 1 2x 10−5
0
2
4
6
8
10
x 105 Pairs (Rpi , Rpj)
Differential current [A]
Freq
uenc
y [#
cou
ples
dev
ices
gen
erat
ing
that
cur
rent
]
Currents distrib. associated to all possible pairs (Ri,Rj)
−2 −1 0 1 2x 10−5
0
2
4
6
8
10
x 105 Pairs (Rapi , Rapj)
Differential current [A]
Freq
uenc
y [#
cou
ples
dev
ices
gen
erat
ing
that
cur
rent
]
Oxide thickness effects on RP/RAP populations
58
• Understand how far to push resistance dispersion of MTJs
• Play with process variation to get a good base for a PUF implementation
• At same time, need reliable memory cells for classic applications: data storage!
Oxide thickness the center of gravity
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
5
10
15
20
25
30
35
40
45
50
STT−MTJs σRap
and σRap
, (% over mean value) − Sample size = 5000
TMR = 120%V
read = 0.1V , Feat.size = 65nm (both w/out variation)
For different oxide thicknesses
σtox
[%]
σR
p ,
σR
ap
[%]
σRp
@ tox
= 0.85nm
σRap
@ tox
= 0.85nm
σRp
@ tox
= 1nm
σRap
@ tox
= 1nm
σRp
@ tox
= 1.15nm
σRap
@ tox
= 1.15nm
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
5
10
15
20
25
30
35
40
45
50
STT−MTJs σRap
and σRap
, (% over mean value) − Sample size = 5000
TMR = 120%V
read = 0.1V , Feat.size = 65nm (both w/ 5% variation)
For different oxide thicknesses
σtox
[%]
σR
p ,
σR
ap
[%]
σRp
@ tox
= 0.85nm
σRap
@ tox
= 0.85nm
σRp
@ tox
= 1nm
σRap
@ tox
= 1nm
σRp
@ tox
= 1.15nm
σRap
@ tox
= 1.15nm
Results validation (PART II)
59
• Strong influence of tOX on RP/RAP: makes sense (e- tunneling)
• No variability whatsoever à back to ideal case (nom. RP/RAP)
• tOX ↑ and σtox ↑à RP/RAP dispersion ↑
• tox plays a dominant role, but when lower than 1nm also
variability of other physical quantities matters
• Digital application: need separation between
the two populations
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5−14
−12
−10
−8
−6
−4
−2
0
2x 10
4
STT−MTJs Rp and Rap populations DISTANCE amount − Sample size = 5000TMR = 120%
Vread
= 0.1V , Feat.size = 65nm (both w/ 5% variation)
For different oxide thicknesses
σtox
[%]
Dis
tance
betw
een p
opula
tions
min
(Rap)
− m
ax(
Rp)
[Ω]
min(Rap)−max(Rp) @ tox
= 0.85nm
min(Rap)−max(Rp) @ tox
= 1nm
min(Rap)−max(Rp) @ tox
= 1.15nm
Cases 1-2-3-4 memory size comparison
60
• Trend for MRAM to store data: differential approach à 2 MTJ +2 MOS for cell
• New schemes (3&4) use differential cells (2T2MTJ)
• Memory size?
PUF Quality: Equations
61
• UNIQUENESS: expresses randomness of the signature, how unique it is w.r.t. others
o Hamming Distance
o Hamming Weight
o Bit Aliasing
• RELIABILITY: stability over undesired parameter fluctuations (voltage, temperature)
o BER: Bit-Error Rate, related to sense amplifier quality and cells reading voltage
Uniqueness & Reliability: example for Case 4
62
44 46 48 50 52 54 560
200
400
600
800
1000
1200
1400
UNIQUENESS: Inter−die Hamming Distance distributionCASE 2T2MTJ+2T2MTJ, NO REFERENCE
Average(HD)=49.995916% ; Deviation(HD)=1.573119%
HD [%]
Count [#
]
10 20 30 40 50 60 70 80 90 100
40
45
50
55
60
Hamming WeightCASE 2T2MTJ+2T2MTJ, NO REFERENCE
Average: 50.12% ; Deviation: 1.40%
CHIP#H
am
min
g W
eig
ht [%
]
Hamming Weight of each chipAverage Hamming Weight
100 200 300 400 500 600 700 800 900 1000
30
35
40
45
50
55
60
65
70
75
Bit AliasingCASE 2T2MTJ+2T2MTJ, NO REFERENCE
Average: 50.12% ; Deviation: 5.02%
Bit Position in each CHIP matrix (reorganized by rows)
Bit
Alis
ing [%
]
Bit Alising of each chipAverage Bit Alising
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 10−6
0
10
20
30
40
50
60
70
80
Sense amplifier min ∆I detectable [A]
Bit
err
ors
ra
te [
%]
Bit errors rate in (2T2MTJ + 2T2MTJ) config.Couples AntiPar + Par.
min ∆ I range: (0 − 5)µA tox = 0.85 nm − Feat.size = 65 nm − TMR = 120 %
Reversed Combo (AP+P)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 10−8
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Sense amplifier min ∆I detectable [A]
Bit
err
ors
ra
te [
%]
Bit errors rate in (2T2MTJ + 2T2MTJ) config.Couples AntiPar + Par.
min ∆ I range: (0 − 0.05)µA tox = 0.85 nm − Feat.size = 65 nm − TMR = 120 %
Reversed Combo (AP+P)
Challenges space Ω
63
• The total number of different challenges (combinations)
• Strong PUF à # of possible challenges exponentially increasing with memory size
0 5 10 15 20 25 30 3510
0
105
1010
1015
1020
1025
1030
1035
1040
Growth rate of the challenges space Ω as function of the word size L. Fix M=32 words
Memory matrix word width L
Challe
nges
space
Ω
• L: # elementary memory cells per row
• M: memory length (# rows)
• K: # cells grouped to each sense amplifier input
• S: # available sense amplifiersA
SEM
ILOG-
Y PL
OT!
Response space Λ
64
• The number of possible responses depending on the selected challenge
• Sense amplifiers output: either 0 or 1
• S/2 sense amplifiers, M memory rows
0 5 10 15 20 25 30 350
0.5
1
1.5
2
2.5x 10
6
Growth rate of the size of the PUF response as function of the word size L. Fix M=32 words
Memory matrix word width L
Siz
e Λ
of th
e b
inary
resp
onse
R