Slide 1 of 47

1
Project IEEE P802.15 Working Group for Wireless
Personal Area Networks (WPANs) Submission Title
The ParthusCeva Ultra Wideband PHY
proposal Date Submitted 05 May,
2003 Source Michael Mc Laughlin, Vincent
Ashe Company ParthusCeva Inc. Address 32-34
Harcourt Street, Dublin 2, Ireland. Voice353-1
-402-5809, FAX -, E-Mailmichael.mclaughlin_at_p
arthusceva.com Re IEEE P802.15 Alternate PHY
Call For Proposals. 17 Jan 2003 Abstract Propos
al for a 802.15.3a PHY Purpose To allow the
Task Group to evaluate the PHY proposed Notice T
his document has been prepared to assist the IEEE
P802.15. It is offered as a basis for discussion
and is not binding on the contributing
individual(s) or organization(s). The material in
this document is subject to change in form and
content after further study. The contributor(s)
reserve(s) the right to add, amend or withdraw
material contained herein. Release The
contributor acknowledges and accepts that this
contribution becomes the property of IEEE and may
be made publicly available by P802.15.
2
The ParthusCeva PHYProposal

3
Overview of Presentation

• Coding
• DSSS Coding scheme - biorthogonal coding
• Ternary spreading codes
• Reed Solomon FEC code
• Optionally concatenated with convolutional code
• Preamble
• Implementation Overview
• Performance
• Link margin
• Test results
• Throughput, Multiple piconet performance
• Complexity

4
Symbol coding

• 64 biorthogonal signals Proakis1
• 64 signals from 32 orthogonal sequences
• Ternary sequences chosen for their
  auto-correlation properties
• Code constructed from binary Golay-Hadamard
  sequences

5
Spreading code comparison

• Length 32 code chosen for aamf and best matching
  with bit rates.

6
Sample rate and pulse repetition frequency

• Signal bandwidth chosen is 3.85GHz to 7.7GHz
• Sampling rate chosen is 7.7Ghz
• 32 chips per codeword, 6 channel bits / symbol

7
FEC Scheme

• Concatenated code for 110 , 220Mbps, 880Mbps
• Reed Solomon outer code (235,255)
• Convolutional rate 4/6 inner code
• Reed Solomon code (43,63) for 490Mbps, 980Mbps

8
FEC scheme - inner code

• A 0.667 rate (rate 4/6) convolutional code was
  chosen for the inner code at 110 and 200 Mbps.
  Proakis2
• Very low complexity 16 state code, constraint
  length 2, Octal generators 27, 75, 72.
• Each of 16 states can transition to any other
  state, outputting 16 of 64 possible codewords.
• Provides 3dB of gain over uncoded errors at a
  cost of 50 higher bit rate

9
Preamble Sequence
10
PAC properties

• Because of the perfect autocorrelation property,
  the channel impulse response can be obtained in
  the receiver by correlating with the sequence and
  averaging the results.
• Because the sequence consists of mostly 1, -1
  with a small number of zeros, correlation can be
  economically implemented. (a length 553 PAC has
  24 0s)

11
Preamble properties

• Very good detect rate and false alarm
  probability. Pfa and Pmd lt 10-4 for CM1 to CM4
  test suite at 10 metres. Detected in 2?s using
  matched filter architecture.
• Matched filter is equivalent of 553 parallel
  correlators
• Different length sequences means other piconets
  wont trigger detection i.e. Pfa still lt 10-3 for
  a different piconets PACn, even at 0.3m
  separation.
• Preamble length varies from 5?s to 15?s
  depending on the bit rate. Lower bit rates use
  longer preambles (Longer distances need more
  training time)

12
PHY Header

• The PHY header is sent at an uncoded 45Mbps rate,
  but with no convolutional coding. It is repeated
  3 times.
• The PHY header contents are the same as 802.15.3
  i.e. Two octets with the Data rate, number of
  payload bits and scrambler seed.

13
Scrambler/Descrambler

• It is proposed that the PHY uses the same
  scrambler and descrambler as used by IEEE
  802.15.3

14
Typical Tx/Rx configuration
Antenna
Single Chip Possible
Output data at 55 - 960 Mbps
Channel Matched filter (Rake Receiver)
Fine/ Band Reject Filter
Band Pass Filter
A/D 7.7GHz, 1 bit
Viterbi Decoder
Correlator Bank
Switch / Hybrid
Descramble
LNA
8-240M symbols/sec
256 - 3800 Mchips/sec
Input data at 55- 960 Mbps
Band Pass Filter
Code Generator
Chip to Pulse Generator
Convolutional encoder
Band Reject Filter
Scramble
Can be avoided with good LNA dynamic range
15
Possible RF front end configuration

• Total Noise Figure 6.6dB

NF 0.3dB (input referred)
NF 2.0dB
NF 3.5dB
BP Filter
Fine Filter
LNA
To Rx
NF 0.8dB
Tx/Rx switch / hybrid
Band reject Filter
From Tx
Depending on Local National or User
requirements
Can be avoided with good LNA dynamic range
16
Matched Filter configuration
Cn
Di
CnN
Di-N
4
1
Cn1
Di-1
Di-N-1
4
1
CnN1
..
..
4x
4x
4x
4x
4
4
4
4
..
4 bit adder
4x
4x

5 bit adder

..
..
17
Matched Filter configuration

• Structure repeated 16 times e.g. a 500 tap filter
  with 4 bit coefficients would have 500 x 16 x 4
  AND gates in first stage
• Calculates 16 outputs in parallel, each runs at
  (480/mps) MHz.
• e.g. 120MHz for 220Mbps
• Multiplier is 4 AND gates.
• First adder stage is 4 OR gates. Very little
  performance loss. (0dB for CM1-3, 0.23dB for
  CM4).
• Coefficients are pre-processed to remove smallest
  if two clash.
• mps is max pulses/sample. 1440/(channel bit
  rate (Mbps))

18
Matched filter

• 560 tap filter takes 135k gates or 0.82 sq mm in
  0.13? standard cell CMOS
• Worst case power consumption 120mW ( at 490Mbps
  ), proportional to data rate. Much lower for CM1
  because of fewer taps.
• Matched filter re-used for correlation with
  training sequence during training phase
• All simulations were carried out with this
  filter/correlator structure

19
(No Transcript)
20
Distance achieved for mean packet error rate of
best 90 8
21
Distance achieved for at worst packet error rate
of best 90 8
22
110 Mbps average PER
23
220 Mbps average PER
24
490 Mbps average PER
25
Multiple Piconet Interferers

• Tests were done according to the Multiple Piconet
  interference procedure outlined in the latest
  revision of the selection criteria (03031r11).
• The distance to the receiver under test was set
  at 0.707 of the 90 link success probability
  distance.
• Tests results were obtained for 1,2 and 3
  interfering piconets

26
Single adjacent piconet
27
Two adjacent piconets
28
Three adjacent piconets
29
Co-channel interference

• Different piconets use exactly the same data mode
  codes as each other.
• Separation is achieved because
• a) a different piconet will have a different
  impulse response and thus will not correlate with
  the matched filter which has been trained for the
  piconet of interest.
• b) Codes wont be synchronised
• Co-channel data mode interference is exactly the
  same as adjacent channel interference.
• Training to the preamble will be affected more
  markedly by co-channel interference. Difficult to
  simulate.

30
Co existence

• Out of band signals, e.g. 802.11b, (lt 3.1GHz and
  gt10.6GHz) are always filtered out.
• Any desired in band energy can be filtered out,
  with minimal effect on performance because the
  whole band is used to transfer data.
• Only adverse effect is the transmit power
  reduction (e.g. Dropping 400MHz for 802.11a loses
  lt0.5dB)

31
Co-existence with 802.11a

• Filtering out the UNII band spectrum from the
  transmitter has very little effect on the
  performance
• The receive matched filter will cope with it
  automatically
• Only 1.25dB of power is lost by filtering out the
  Tx signal from 5GHz to6GHz
• This is the equivalent of a 15 loss in distance

32
Interference and susceptibility

• As for co existence, out of band signal, e.g.
  802.11b, are always filtered out.
• Again, any desired in band energy can be filtered
  out, with minimal effect on performance because
  the whole band is used to transfer data.
• Only adverse effect is the receive power
  reduction (e.g. Dropping 400MHz for 802.11a loses
  lt0.5dB), its just a part of the channel.

33
Narrowband interference

• Immunity to narrowband interference
• With no filtering
• Processing a gain of e.g. 24dBs at 110Mbps. Any
  interfering tone is reduced by this amount.
• With digital notch filter
• Tones can be detected at the A/D output.
• A simple notch filter either at the input or
  output of the matched filter can then remove this
  completely with no loss in performance (if notch
  is narrow enough)

34
PHY-SAP Data Throughput

• At higher bit rates, a 1024 byte frame is very
  short.
• The channel will be stationery for more than one
  frame so it is possible to send multiple frames
  for each preamble.
• T_MIFS1µs, T_SIFS5µs, T_PHYHDR1.1µs,T_HCS0.29µ
  s, T_MACHDR1.45µs

35
Scalable solution

• 55 - 980 Mbps. Gate count depends on maximum bit
  rate and power consumption of baseband PHY is
  proportional to bit rate.
• 880Mbps has 90 link success on CM1-CM3 at over
  3.4m, 2.8m and 2.6m with exactly same RF and
  sample rate as the other rates
• Receiver here uses 3.85 - 7.7GHz.
• 2.5dB extra performance gain if full band used.
• 1.0dB lower performance if 3.2-4.8 GHz band used
  with 50 power reduction

36
Complexity - Area/Gate count, Power consumption

• These figure are for a standard cell library
  implementation in 0.13µm CMOS

37
How can it be so good?

• Where does this large Performance/Cost/Power
  advantage come from
• Compare with two proposals other proposals - TFI
  OFDM and Multiband
• These two were chosen because of their prominence
  and fairly comprehensive results available

38
Performance- How much better?

• All similar under AWGN, i.e. 20 odd metres at
  110Mbps.
• 200/220/240 Mbps
• 50 farther than either TFI-OFDM or Multiband
  over CM1
• 92 farther than TI-OFDM and 230 farther than
  Multiband over CM4
• 110/120 Mbps
• 22/25 farther over CM1
• 22 farther than TFI-OFDM and 75 farther than
  Multiband over CM4
• 480 Mbps - 86 farther than TFI-OFDM over CM1, no
  CM4 or Multiband figures available.
• NB The distances quoted for TFI-OFDM and
  Multiband do not take into account the 4.7dB loss
  required for FCC compliance i.e. a further 1.72
  gain factor.

39
Performance - Why is it better?

• Gap is small with no multipath at low rates,
  larger as the multipath increases and speed
  increases.
• Multiband approach only gathers a small amount of
  multipath energy. 16ns at 120Mbps and 8ns at
  240Mbps. CM4 channels have significant energy
  spread over 100ns
• ParthusCeva PHY - has equivalent of a 230 finger
  rake.
• Ternary codes were chosen for their multipath
  immunity

40
Power

• 110/120Mbps Rx _at_ 0.13
• This approach up to 145mW
• Multiband approach 170-200mW
• TFI - OFDM 205mW
• Why so good?
• Simple analog Rx section
• Single bit ADC
• No AGC
• No mixer
• No FFT in the receiver. Matched filter with 1 bit
  inputs. Low complexity decoders

41
Cost

• Comparison at 0.13µ
• This approach 4.25mm2
• Multiband approach 7.3mm2
• TFI - OFDM approach 6.9mm2
• Why the difference ?
• Same reasons as power, low complexity digital and
  analog requirements

42
Summary of advantages

• Ternary spreading codes
• Better auto-correlation properties
• Perfect PAC training sequence
• Simple RF section
• 1 bit A/D converter
• No AGC required
• No Rx mixers required
• Long rake possible - near multipath immunity
• 4 bit coefficients
• 1 bit data
• no multipliers
• Cost and Power very similar to Bluetooth

Low cost
Low power consumption
Low Noise figure
43
The ParthusCeva PHY
Faster
44
The ParthusCeva PHY
Faster
Smaller
45
The ParthusCeva PHY
Faster
Smaller
Cooler
46
Backup Slides
47
Self evaluation General Criteria
48
Self evaluation PHY protocol
49
Self evaluation MAC enhancements
50
Ternary orthogonal sequences

• From any base set of 32 orthogonal binary
  signals, can generate 32C16 sets of 32
  orthogonal ternary sequences.
• Generate by adding and subtracting any 16 pairs.
• Generally, if the base set has good correlation
  properties, so will a generated set.

51
Good base binary set

• Base set is a set of binary Golay-Hadamard
  sequences
• Take a binary Golay complementary pair.
• s1161 1 1 1 1 1 -1 -1 -1 1 1 -1 -1 1 -1 1
• s2161 1 1 1 -1 -1 1 1 -1 1 1 -1 1 -1 1 -1
• if Acirculant(s116) and Bcirculant(s216)
• and G32 A B
• BT -AT
• then G32 is a Hadamard matrix. Seberry
• This type has particularly good correlation
  propertiesSeberry
• Detector can use the Fast Hadamard Transform

52
Creating Orthogonal Ternary Sequences

• Take a matrix of binary orthogonal sequences
• Add any two rows to get a ternary sequence
• Sum of any other two rows is orthogonal to this
• Continue till all the rows are used
• Repeat but subtracting instead of adding

53
Orthogonal Ternary Example

• E.g. 1 1 1 1
• 1 -1 1 -1
• 1 -1 -1 1
• 1 1 -1 -1
• pairing 1 with 3 and 2 with 4 gives this
  orthogonal matrix
• 2 0 0 2
• 2 0 0 -2
• 0 2 2 0
• 0 -2 2 0

54
Finding good Ternary Golay Hadarmard codes

• Large superset of orthogonal sequence sets to
  test
• Define aperiodic autocorrelation merit factor
  (aamf) as the ratio of the peak power of the
  autocorrelation function to the mean power of the
  offpeak values divided by the length of the code.
• Random walk used to find set with best aamf

55
Code comparison

• Length 32 code chosen for aamf and best matching
  with bit rates.

56
Rate 4/6 Convolutional coder

3 bits out
Map every 6 bits to one of 64 biorthogonal
codewords

1 of 64

2 bits in
57
Ternary Orthogonal Length 32 Code Set

• 0 - 0 - 0 - 0 0 0 - 0 0 0 - 0 - 0 - 0
  - 0 - 0 0 - 0
• - 0 - 0 - 0 - 0 0 0 0 0 0 0 0 0 - 0 0
  0 - 0 - - - -
• 0 0 0 0 - - 0 0 0 0 0 0 0 0 - 0 0 - - -
  - 0 0 - - -
• 0 0 0 - - 0 0 - 0 0 0 - 0 0 0 0 - - 0
  0 - 0 - - 0 -
• - 0 0 0 0 - - 0 - 0 0 0 - - 0 0 0 -
  - 0 - 0 0 - 0 0
• 0 0 0 - 0 0 0 0 0 0 - 0 0 0 - 0 0 - -
  - 0 0 - - - -
• 0 - 0 0 0 - - 0 0 - 0 0 0 - 0 0 0 0
  - - 0 0 - 0 - -
• 0 0 0 0 - 0 - - - 0 0 0 0 0 0 0 - 0 0 0
  - - - 0 - -
• 0 0 0 0 0 0 - 0 0 0 0 0 0 - - - - - 0 0 -
  - - 0 0 -
• 0 0 0 0 0 0 0 - 0 - - - 0 - - - 0 0 0 0
  0 0 0 - - -
• 0 0 0 0 0 0 0 0 0 0 - - 0 0 - 0 0 - - -
  - 0 0 - - -
• - 0 0 0 0 0 0 0 - 0 - - - 0 - - - 0 0
  0 0 0 0 0 - -
• 0 0 0 0 0 0 0 0 0 0 0 0 - - - - 0 0 - -
  - - 0 0 - -
• - - - 0 0 0 0 0 0 0 - 0 - - - 0 - - -
  0 0 0 0 0 0 0
• 0 0 0 0 - 0 0 0 0 0 0 - 0 0 - - 0 0 - - -
  - - 0 0 -
• 0 - - - 0 0 0 0 0 0 0 - 0 0 - - - 0 -
  - - 0 0 0 0 0
• 0 0 - 0 0 0 0 0 0 - 0 0 - 0 0
  0 - 0 - 0 - 0
• 0 0 0 - 0 0 0 0 0 - - 0 - 0 -
  0 - 0 0 0 0 0
• - 0 0 - - 0 0 - 0 0 0 0 0 0
  0 0 - - 0 0 0 0

58
Matlab Code to generate PAC sequences

• function phiipatov(nu,multiplier,mul2)
• Generate a length nu, ternary perfect periodic
  autocorrelation sequence
• using Singer Cyclic Difference Sets e.g. (553,
  24, 1)
• function phiipatov(nu,multiplier,mul2)
• if nargin1 multiplier1 mul2-1 end
  multipliers 1,-1 are most commonly good
• if nargin2 mul2-1 end
  multipliers 1,-1 are most commonly good
• phi0
• if gcd(nu,multiplier)gt1
  must not be a common divisor of nu
• return
• end
• if gcd(nu,mul2)gt1
  must not be a common divisor of nu
• return

59
References

• Proakis1 John G. Proakis, Digital
  Communications 2nd edition. McGraw Hill. pp
  224-225.
• Proakis2 John G. Proakis, Digital
  Communications 2nd edition. McGraw Hill. pp
  466-470.
• Seberry et al J. Seberry, B.J. Wysocki and T.A.
  Wysocki, Golay Sequences for DS CDMA
  Applications, University of Wollongong
• Ipatov V. P. Ipatov, Ternary sequences with
  ideal autocorrelation properties Radio Eng.
  Electron. Phys., vol. 24, pp. 75-79, Oct. 1979.
• Høholdt et al Tom Høholdt and Jørn Justesen,
  Ternary sequences with Perfect Periodic
  Autocorrelation, IEEE Transactions on
  information theory.