Full Duplex Radios

Sachin Katti

Kumu Networks & Stanford University

4/17/2014 1

“It is generally not possible for radios to receive and transmit

on the same frequency band because of the interference that

results.”

- Andrea Goldsmith, “Wireless Communications,” Cambridge Press, 2005.

Why are radios half duplex?

4/17/2014 2

TX

RX

RX

TX

Radio 1 Radio 2

“It is generally not possible for radios to receive and transmit

on the same frequency band because of the interference that

results.”

- Andrea Goldsmith, “Wireless Communications,” Cambridge Press, 2005.

Why are radios half duplex?

4/17/2014 3

TX

RX

RX

TX

Radio 1 Radio 2

Self-Interference is a hundred billion times

(110dB+) stronger than the received signal

Isn’t this easy to solve?

After all we know the interfering signal, why can’t

we just “subtract” it?

4/17/2014 4

4/17/2014 5

Signal Sent

Do we know what we are transmitting?

Signal Sent T

x

PA

DAC

TX

Mixer

4/17/2014 6

Signal Sent

Do we know what we are transmitting?

Signal Sent T

x

PA

DAC

TX

Centered at Carrier Freq (2.45GHz)

Mixer

4/17/2014 7

Signal Sent Signal Sent T

x

PA

DAC

TX

Centered at Carrier Freq (2.45GHz)

Mixer

Do we know what we are transmitting?

If you were to cancel, how much do we need?

Transmitted Signal

Pow

er

in d

Bm

-10 dBm Harmonics

-90 dBm Receiver Noise floor

Cancel residual

in digital to

reach noise floor

20 dBm Average Power

-20 dBm Transmit noise

4/17/2014 8

Cancel entire 110

dB to reach

noise floor

Cancel 70 dB Tx

noise to reach

noise floor

Cancel 70 dB in

Analog in such a

way to eliminate TX

noise

-80 dBm Harmonics

Cancelled Signal

Takeaways: Require 110dB of total cancellation, of which at

least 70dB has to eliminate transmitter noise in analog.

Contributions

We have invented in-band single

antenna full duplex radios • Self-Interference cancellation that

eliminates everything to the noise floor

• Practically achieves close to expected

theoretical 2x throughput increase

4/17/2014 9

RX RF Frontend

TX RF Frontend

T

+iT

+δT

Circulator

Contributions

We have invented in-band single

antenna full duplex radios • Self-Interference cancellation that

eliminates everything to the noise floor

• Practically achieves close to expected

theoretical 2x throughput increase

Algorithms & circuits to

estimate transceiver distortion

and cancel self interference • Hybrid (analog & digital) design with RF

cancellation circuit and DSP algorithms

4/17/2014 10

RX RF Frontend

TX RF Frontend

T

+iT

+δT

Circulator

Σ Analog Cancellation Eliminates all Tx noise and Protects ADC from saturating

Digital Cancellation Eliminates all Linear and Non-Linear Distortion

(Harmonics)

Σ

Mixed RF/Digital Design: Analog + Digital Cancellation

4/17/2014 11

Circulator (-15dB)

T

T

R+iT T

TX

RF Frontend

RX

RF Frontend

Digital Baseband

Analog RF Cancellation

4/17/2014 12

R

Circulator (-15dB)

T

T

T

TX

RF Frontend

RX

RF Frontend

Digital Baseband

εT iT’

RF Cancellation Circuit

Adaptive Algorithms

R+iT

+δT

4/17/2014 13

a1

a6

a2

a5

a3

a4

a8

a7

TX

RF Frontend

RX

RF Frontend

Adaptive Algorithms

Analog RF Cancellation

d4

d1

d2 d3

d5

d6 d7

d8

εT iT RF Cancellation Circuit

4/17/2014 14

a1

a6

a2

a5

a3

a4

a8

a7

TX

RF Frontend

RX

RF Frontend

Adaptive Algorithms

Analog RF Cancellation

d4

d1

d2 d3

d5

d6 d7

d8

εT iT

Analog Theory Intuition: Branch Delays

time (delay) d1 d2

fixed delays

N delay attenuation branches

control algorithm

Σ

interference signal

Delays are fundamentally related to sampling theory

4/17/2014 15

d1 a1

d2 a2

a3 d3

a4 d4

d

Analog Theory Intuition: Branch Delays

time (delay) d1 d2 d3 d4

fixed delays

N delay attenuation branches

control algorithm

Σ

interference signal

Delays are fundamentally related to sampling theory

4/17/2014 16

d1 a1

d2 a2

a3 d3

a4 d4

d

First branch pair: positive

4/17/2014 17

Estimating Branch Attenuation How do we fix attenuation ranges?

d2 d1 d

First branch pair: positive

4/17/2014 18

Estimating Branch Attenuation How do we fix attenuation ranges?

d2 d1 d

a1 a2

Second branch pair (negative)

Estimating Branch Attenuation How do we fix attenuation ranges?

4/17/2014 19

d4 d3

d2 d1 d

a1 a2

Estimating Branch Attenuation How do we fix attenuation ranges?

4/17/2014 20

Second branch pair (negative) a4

a3

d4 d3

d2 d1 d

a1 a2

Estimating Branch Attenuation Adaptation to environmental changes: Assumption d known

4/17/2014 21

d4 d3

d2 d1 d

a1 a2

a4 a3

Estimating Branch Attenuation Adaptation to environmental changes

4/17/2014 22

d4 d3

d2 d1 d

a1

a2

a4

a3

Estimating Branch Attenuation Adaptation to environmental changes

4/17/2014 23

d4 d3

d2 d1 d

a1

a2

a4

a3

Digital Baseband Cancellation

4/17/2014 24

isolator (-15dB)

T

Σ

T

T

TX

RF Frontend

RX

RF Frontend

Digital Baseband Cancellation Eliminates 2nd+ Order Non-Linearities (e.g. Intermod Products, LO

leakage, IQ imbalance)

RF Cancellation Circuit

Adaptive Algorithms

+δT

+ iT

Σ

Digital Baseband Cancellation

4/17/2014 25

TX

RF Frontend

RX

RF Frontend

Digital Baseband Cancellation Eliminates 2nd+ Order Non-Linearities (e.g. Intermod Products, LO

leakage, IQ imbalance)

RF Cancellation Circuit

Adaptive Algorithms

Digital Self-Interference Cancellation

• Challenge: Need to cancel main signal as well as higher order

harmonics upto the 11th order

• Prior approaches only cancel main signal, ignore hamonics

• Naïve approach to non-linearities: Needs to estimate ~1200

coefficients, would require a large number of training symbols &

hardware resources, infeasible in practice

• Our approach: Compact and fast digital self-interference

cancellation algorithm (needs to only estimate ~200 coefficients,

works with existing WiFi packet format)

4/17/2014 26

Evaluation Q1: Does it work with commodity radios?

• Challenge: Extremely high transmitter noise and non-linearities

• 20MHz BW (transceiver limitation) • 25dBm max TX power • WiFi 802.11n PHY

Goal: Build a full duplex radio using a cheap $2 COTS Maxim transceiver

• Commodity

transceiver

Digital

~40dB + Total

>110dB = Analog

>70dB 4/17/2014 28

-100

-85

-70

-55

-40

-25

-10

5

20

2.43 2.44 2.45 2.46 2.47

Po

we

r in

dB

m

Freq in Ghz

WARP 20 Mhz

Tx Signal

Residual Signalafter AC

Residual Signalafter DC

Noise Floor

72 dB

38 dB

Evaluation Q1: Does it work with commodity radios?

• Commodity

transceiver

• Tunes to

environmental

changes within

8us, needs to

be re-tuned

every 100ms

Digital

~40dB + Total

>110dB = Analog

>70dB 4/17/2014 29

-100

-85

-70

-55

-40

-25

-10

5

20

2.43 2.44 2.45 2.46 2.47

Po

we

r in

dB

m

Freq in Ghz

WARP 20 Mhz

Tx Signal

Residual Signalafter AC

Residual Signalafter DC

Noise Floor

72 dB

38 dB

Evaluation Q1: Does it work with commodity radios?

How do we compare against prior designs? 20 MHz Bandwidth. WiFi OFDM waveform, 25 dBm TX power

4/17/2014 30

Compared Approaches

Our Design

Balun Cancellation (Mobicom’11)

Extra-Tx Chain Design (Sigcomm’11, Asilomar’11)

How do we compare against prior designs? 20 MHz Bandwidth. WiFi OFDM waveform, 25 dBm TX power

4/17/2014 31

Compared Approaches Cancellation in (dB)

Our Design 110

Balun Cancellation (Mobicom’11)

85

Extra-Tx Chain Design (Sigcomm’11, Asilomar’10)

80

How do we compare against prior designs? 20 MHz Bandwidth. WiFi OFDM waveform, 25 dBm TX power

4/17/2014 32

Minimum SNR required for receiving a packet

Self-interference residue over noise floor

>

Compared Approaches Cancellation in (dB)

Self-interference residue over noise floor (dB)

Our Design 110 ~1

Balun Cancellation (Mobicom’11)

85 25

Extra-Tx Chain Design (Sigcomm’11, Asilomar’10)

80 30

Evaluation Q2: Does that translate to doubling of

throughput in practice?

4/17/2014 33

Throughput of FD Throughput of HD

Gain =

• Testbed: Indoor office noisy environment, various locations for the two full duplex radios.

• Compare throughput achieved in full duplex with that achieved in half duplex

• Full duplex implemented using our approach, and prior balun and extra TX chain based approaches

4/17/2014 34

0

0.2

0.4

0.6

0.8

1

0 0.5 1 1.5 2

CD

F

Gain vs Half Duplex

Balun Cancellation Extra Transmitter Our Design

Worse than standard Half Duplex

1.97x median gain

Evaluation Q2: Does that translate to doubling of

throughput in practice?

Our design achieves the theoretical throughput doubling

Self-interference cancellation is broadly applicable

Full Duplex Radios

4/17/2014 35

Transmitted Self-Interference

Proprietary and Confidential

Channel 1 Channel 2

In-Band Full

Duplex

Double capacity

Received Signal

Adaptive

Frequency Division

Full Duplex (FDD)

Flexibly decide which

channels to transmit &

receive on

Transmitted Self-Interference

Received Signal

Channel 1 Channel 2

Applicable To a Host of Problems

LTE Access High performance Relay

Node. Doubles spectral

efficiency for TD-LTE.

WiFi Access Dense coverage by avoiding

interference between

adjacent bands

Mobile Devices World phones supporting

any FDD channel pairs with

adaptive duplexers

PtP and PtMP

Backhaul Doubles spectral efficiency,

and mitigate interference in

unlicensed bands

Full Duplex Everywhere

36 Proprietary and Confidential

To Conclude

Key contribution: Cancellation design that eliminates all self-

interference to the noise floor

– Full duplex radio is one application of this interference

cancellation technique.

– Widely applicable (Picasso, IMDShield, WiVi, Dhwani, …)

Emphasizes the need for an interdisciplinary approach that

combines RF circuit design, signal processing and

communication algorithm design

4/17/2014 37