Montgomerys Multiplication Technique: How to make it Smaller and Faster

1
Montgomerys Multiplication Technique How to
make it Smaller and Faster

• Colin D. Walter
• Computation Department, UMIST, UK
• www.co.umist.ac.uk

2
Peter Montgomery

• Modular Multiplication without Trial Division
  Math. Computation, vol. 44 (1985)
  519-521
• (A ? B) mod M
  without obtaining digits q ? (A ? B) / M

3
Motivation

• Faster RSA Cryptosystem
• ? through pipelined array
• Safer encryption
• ? against timing or DPA attacks

4
Overview

• RSA Notation
• Classical Algorithm
• Montgomerys Version
• Comparison
• carry propagation
• digit distribution
• communication
• timing/power attacks
• Conclusion

5
Enigma

• Special Purpose Colossus (1943-44)
• Tommy Flowers,
• Bletchley Park, England.
• General Purpose ENIAC (1943-46)
• John Eckert John Mauchly
• Philadelphia, US.

6
RSA

• Modulus M of around 1024 bits
• Two keys d and e such that Ade ? A mod M
• A encrypted to C Ae mod M
• C decrypted by A Cd mod M
• M PQ, a product of two large primes
• e is often small (e.g. a Fermat prime)
• d satisfies de ? 1 mod (P1)(Q1)

7
Faster H/W More Secure Encryption

• Work to factorize M doubles for every
  extra 15 bits (for key lengths 210
  bits)
• Work to en/decrypt
• ((102415)/1024)2 per multiplication
• ((102415)/1024)3 per exponentiation,
• i.e. only 5 extra!

8
Number representations
n1

• X ?i0 xiri
• r 2k is the radix (prime to M)
• xi is the ith digit (usually 0 ? xi lt r)
• n ? max no. of digits in any number
• Redundant reps
• wider digit range than 0 .. r?1
• H/W is built from k?k-bit multipliers
• n fixed by H/W register size

9
Redundancy

• Digits xj split into carry-save parts xj xj,s
  rxj,c
• X ? XY
• is performed by digit-parallel addition
• xj ? xj,s xj?1,c yj
• No carry propagation only old carries on right
  side

10
Multiplication A?B

• Use n digit multipliers to form ai?B and add to a
  partial product P
• P 0
• For i n?1 downto 0 do
• P r?P ai?B
• Post-condition P A?B

11

• Either Use redundancy in P and parallel digit
  addition to add aiB in one clock cycle
• Cell j computes aibj in cycle i

ai
ai
pj1,s
pj,c
pj,s
pj-1,c
pj-1,s
bj1
bj
bj-1
cell j
cell j-1
cell j1
pj,c
pj1,c
pj1,s
pj,s
pj-1,c
pj-1,s
P P ai?B (digit-parallel)
P in Carry-Save form pj pj,s r?pj,c
12

• or Pipeline the addition of ai?B over n cycles
  and propagate carries with no redundancy
• Cell j computes aibj in cycle ij

pj1 bj1
pj bj
pj-1 bj-1
ai
ai
ai
ai
time j1
time j
time j-1
carry
carry
carry
carry
pj1
pj
pj-1
P P ai?B (digit-serial)
13
Multiplier Complexity

• Assume wires take area but not time (or power).
• Area?Time2 complexity for un-pipelined k-bit
  multiplication is bounded below by k2
• This can be achieved for time in log k ..?k
• Discrete Fourier Transform has large constants
  for time and area.
• Better, but asymptotically poorer designs for k
  expected here.

14

• Cross-over point ?
• 107 transistors available for RSA ?
• k ? 64 to accommodate ai?B
• Speed by using at least n multipliers to perform
  a full length ai?B (or equivalent) in one cycle.
  

15
Real-Time ?

• Assume
• bus is one k-bit digit per cycle
• k-bit multiplier operates in one cycle
• Then
• A?B takes n cycles using n multipliers
• Throughput is one digit per cycle for multn.
• Need O(nk) multiplications for decryption
• Conclude
• Need O(nk) rows of n multipliers.

16
Classical Mod Multn Algorithm

• Pre-condition 0 ? A lt rn
• P 0
• For i n?1 downto 0 do
• Begin
• P rP aiB
• qi P div M
• P P ? qiM
• End
• Post-conditions P AB ? QM,
• P ? (AB) mod M

17
Comments

• Carry propagation a problem
• (it slows finding q)
• Use only top digits of M and P to determine
  a good multiple of M to remove
• P is bounded by small multiple of M
• Clean up only at end
• Critical path is finding q.

18
Disadvantages

• Redundant rep. for digit-parallel operation
• Global broadcast of q to each digit position

19
Montgomerys Mod Multn Algm

• Pre-condition 0 ? A lt rn
• P 0
• For i 0 to n?1 do
• Begin
• qi (p0aib0)(-m0-1) mod r
• P (P aiB qiM) div r
• Invariant 0 ? P lt MB
• End
• Post-condition Prn AB QM ,
• P ? (ABrn) mod M

20
Peter Montgomery

• reverses the multiplication order
• chooses digits from least to most significant
• shifts down on each iteration.
• uses the least significant digits
  to determine multiple of M to
  subtract.
• Computes (ABrn) mod M

21

• The factor rn is cleared up in post-processing
• Any extra multiple of M is removed then
• qi has no carries to wait for
• Pipelining of the digits can now take place
• compute aibj1 on the cycle after aibj
• use a non-redundant representation
• no broadcasting of qi

22
The Post-Condition

• m0?1 exists
• qi chosen so division by r is exact
• Define Ai ? j0 ajrj and Qi analogously
• Then Ai Ai?1riai and An A
• So ri1P AiB QiM at end of ith
  iteration
• Hence rnP AB QM at end.

i
23
The Bounds

• A converted on-line to non-redundant form
• Can assume ai ? r?1
• So loop invariant P lt MB

24

• If critical path length is computing q
• Scale M to ensure (?m0?1) mod r 1
• Shift B up to make b0 0
• Result
• qi p0 mod r is simple
• Critical path in repeated cell.
• Cost
• Increase n by 2

25
Removing rn

• The Montgomery class of A is
• A ? rnA mod M
• Montgomery modr multn is denoted ? .
• Montgomery product of A and B is
• A ? B ? A B r?n ? ABrn ? AB mod M.
• Applying ? to A instead of ? to A produces
• Ae in an expn algorithm

_
_
_
_
_
_
_
_
___
_
_
_
___
26
Encryption Process
_
__

• Process A ? A ? Ae ? Ae
• Precompute R2 rn ? r2n mod M
• Start with A ? R2 ? Arn ? A mod M
• Exponentiate to obtain Ae
• End with Ae ? 1 ? Ae mod M

__
_
_
__
__
_
27
2M Bound

• Outputs are re-used as inputs.
• So need to bound I/O
• Suppose an?1 0
• Then P lt MB at end of loop n?2
• yields P lt Mr?1B at very end.
• e.g. If B lt 2M then P lt 2M

28

• Suppose 2rM lt rn, A lt 2M and R2 lt 2M
• Then A lt 2M, Ae lt 2M and P Ae ? 1 lt 2M
• Final output P satisfies
• Prn Ae QM where Q ? rn?1.
• Here Ae lt 2M yields Prn lt (rn1)M So P ? M
• P M ? Ae ? 0 mod M ? A ? 0 mod M
• A M should never arise A 0 yields P 0.
• So no final modular adjustment is necessary.

_
__
__
_
__
__
__
29
Digit-Parallel Implementation

• Classical vs Montgomery
• Similarities
• Broadcasting of qi and ai
• Redundant representations
• Computing qi takes time
• Differences
• Bits to determine qi

30
ai, qi
ai, qi
mj bj
mj-1 bj-1
mj1 bj1
pj,c
pj1,s
pj-1,c
pj-1,s
pj,s
cell j
cell j-1
cell j1
pj1,c
pj,c
pj-1,c
pj1,s
pj-1,s
pj,s
P P ai?B (digit-parallel, not modular)
P in Carry-Save form pj pj,s r?pj,c
31
ai qi
ai qi
mj1 bj1
mn-1 bn-1
mj-1 bj-1
mj bj
qi
pj,s
pj1,s
pn-2,s
pj-1,s
pj-2,s
j
j1
j-1
n-1
qi1
pj-1,c
pj,c
pn-3,c
pj1,c
pj-3,c
pj,c
pj-2,c
Digit-Parallel P rP ai?B - qi?M
(Classical)
32
mi1
mi
bj1
bj
mi-1
m0
b0
bj-1
qi
qi
qi
qi
qi
ai
ai
ai
ai
ai
ai
j
j-1
j1
0
ci,j2
ci,j
ci,1
ci,j1
ci,j-1
(i)
(i)
(i)
(i1)
(i1)
(i1)
pj-1
pj
p0
pj-1
pj
pj-2
(i)
pj1
(n)
(n)
(n)
pj-2
pj-1
pj
Data Flow for P(i1) (P(i) ai?B qi?M)/r
(Montgomery)
33
Systolic Array (Montgomery)

• Write ith value of P as P(i) ? j0 p(i?1) r j
• Cells in col j compute p(i)j at time 2ij
• p(i)j rc(i)j ? p(i?1)j1 c(i)j?1 aibj
  qimj
• Cells in col 0 compute qi at time 2i
• qi ? (p(i?1)1aib0)(?m0?1) mod r
• Any number of rows may be constructed
• Different timing schedules are possible

n?1
34
Systolic Array for P (A?B Q?M)r-n
p(i)
p(i)
p(i)
p(i)
j
j-2
mj-1 bj-1
mj bj
mj1 bj1
j-1
j1
ai
ai
ai
ai
cell i,j1
cell i,j
cell i,j-1
qi
qi
qi
qi
carry
carry
carry
carry
p(i1)
p(i1)
p(i1)
p(i1)
j-1
j-2
j1
j
mj bj
mj1 bj1
mj-1 bj-1
ai1
ai1
ai1
ai1
cell i1,j1
cell i1,j
cell i1,j-1
qi1
qi1
qi1
qi1
carry
carry
carry
carry
mj-1 bj-1
mj bj
mj1 bj1
p(i2)
p(i2)
p(i2)
p(i2)
j-1
j
j-2
j1
35
mi1
mi
bj1
bj
mi-1
m0
b0
bj-1
qi
qi
qi
qi
qi
ai
ai
ai
ai
ai
ai
j
j-1
j1
0
ci,j2
ci,j
ci,1
ci,j1
ci,j-1
(i)
(i)
(i)
(i1)
(i1)
(i1)
pj-1
pj
p0
pj-1
pj
pj-2
(i)
pj1
(n)
(n)
(n)
pj-2
pj-1
pj
Data Flow for P(i1) (P(i) ai?B qi?M)/r
36
Digit-Serial Implementation (Montgomery)

• Advantages
• Local communication
• Shorter critical path
• Critical path easily in repeated cell
• Non-redundant representation
• Digit serial I/O
• Different digits qi and ai re DPA

37
Digit-Serial Implementation (Montgomery)

• Disadvantage
• H/W only half used
• Solutions
• Interleave two multiplications
• E.g. configure exponentiation ? 75 use
• Group digits as per Peter Kornerup 94

38

• Other cell boundaries/groupings are possible
• Timing front angles in the data dependency graph
  can be altered
• For current speed of array implementations see
  Blum and Paar 99
• Vuillemin et al. 97 constructed an array
• Design is parametrised by k and no. of rows.

39
Data Dependency Diagrams
40
Data Dependency Diagrams
Parallel Digit Implementation
t 0
t 1
t 2
t 3
41
Data Dependency Diagrams
Walter 93
t4
t3
t2
t1
t0
t5
1 tick
t6
t7
2 ticks
...
42
Data Dependency Diagrams
Kornerup 94
t0
t1
t2
t3
t4
t5
t6
43
Data Integrity

• P AB ? QM or Prn AB QM
• These are easily checked mod m.
• e.g. m a prime just above the maximum cell
  output.
• Cost one cell in the array i.e. increasing n
  by 1.
• On error, abort or re-compute by another route
• e.g. M replaced by dM for a digit d prime to r.

44
Timing Power Attacks

• Most attacks which succeed on the classical
  algorithm have equivalents which will succeed on
  corresponding implementation of Montgomerys
  algorithm.
• With parallel digit processing, the same digits
  of A and Q are used in every digit slice in the
  same cycle. So DPA might reveal them.
• Pipelined version has no equivalent (see data
  dependency graph). It uses many different digits
  of A and Q in each cycle. DPA is more difficult.

45
Conclusions

• For single k-bit multiplier or array of n
  parallel cells, classical and Montgomery
  algorithms are almost equal.
• For pipelined array, Montgomery method has
  advantages smaller time area constants, better
  I/O, better against DPA
• Pipeline is more complex for 100 use, but faster
  clock.
• Parameters can be chosen for specific purposes.

46

• Go forth and Multiply