Full Duplex MIMO Radios Dinesh Bharadia and Sachin Katti Stanford University 4/25/2014 1
Refresher: Why was full duplex considered impossible? 4/25/2014 2
Refresher: Why was full duplex considered impossible? Radio 1 Radio 2 X RX RX X 4/25/2014 2
Refresher: Why was full duplex considered impossible? Radio 1 Radio 2 X RX RX X 4/25/2014 2
Refresher: Why was full duplex considered impossible? Radio 1 Radio 2 X RX RX X 4/25/2014 2
Refresher: Why was full duplex considered impossible? Radio 1 Radio 2 X RX RX X Self-Interference is a hundred billion times (110dB+) stronger than the received signal 4/25/2014 2
State of the Art: SISO Full duplex (Sigcomm 13) Demonstrated that practical inband SISO (single antenna) full duplex radios are possible Self-Interference cancellation that eliminates everything to the noise floor Practically achieves close to expected theoretical 2x throughput increase Analog Eliminates all x noise and Protects ADC from saturating + +δ X RF Frontend RX RF Frontend Digital Eliminates all Linear and Non- Linear Distortion (armonics) 4/25/2014 3
Why care about full duplex? Full Duplex provides a step jump in Spectral Efficiency 4/25/2014 4
Why care about full duplex? Full Duplex provides a step jump in Spectral Efficiency LE 2X2 1.42 Spectral Efficiency 4/25/2014 4
Why care about full duplex? Full Duplex provides a step jump in Spectral Efficiency LE 2X2 1.42 Spectral Efficiency LE 2X2 Spectral Efficiency is 1.42 bits/sec/z Spectra Efficiency Doubled over last decade 4/25/2014 4
Why care about full duplex? Full Duplex provides a step jump in Spectral Efficiency LE 2X2 1.42 Spectral Efficiency LE 2X2 Spectral Efficiency is 1.42 bits/sec/z Spectra Efficiency Doubled over last decade 4/25/2014 4
Why care about full duplex? Full Duplex provides a step jump in Spectral Efficiency LE 2X2 1.42 Spectral Efficiency LE 2X2 Spectral Efficiency is 1.42 bits/sec/z Spectra Efficiency Doubled over last decade Full Duplex provides a step jump 4/25/2014 4
Why care about full duplex? Full Duplex provides a step jump in Spectral Efficiency LE 2X2 1.42 Spectral Efficiency LE 2X2 Spectral Efficiency is 1.42 bits/sec/z Spectra Efficiency Doubled over last decade Full Duplex provides a step jump Spectral efficiency gains over last decade mostly from MIMO For full duplex to be viable, need to work with MIMO 4/25/2014 4
his Work 4/25/2014
his Work Designed & implemented a near-ideal MIMO in-band full duplex radio 4/25/2014
his Work Designed & implemented a near-ideal MIMO in-band full duplex radio ybrid (analog & digital) cross and self-talk cancellation circuits & algorithms that completely eliminate interference to the noise floor 4/25/2014
his Work Designed & implemented a near-ideal MIMO in-band full duplex radio ybrid (analog & digital) cross and self-talk cancellation circuits & algorithms that completely eliminate interference to the noise floor Low complexity cancellation design 4/25/2014
his Work Designed & implemented a near-ideal MIMO in-band full duplex radio ybrid (analog & digital) cross and self-talk cancellation circuits & algorithms that completely eliminate interference to the noise floor Low complexity cancellation design Prototype 3x3 MIMO full duplex radios using off-the-shelf WiFi radios 4/25/2014
his Work Designed & implemented a near-ideal MIMO in-band full duplex radio ybrid (analog & digital) cross and self-talk cancellation circuits & algorithms that completely eliminate interference to the noise floor Low complexity cancellation design Prototype 3x3 MIMO full duplex radios using off-the-shelf WiFi radios Experimental indoor evaluation which demonstrates that our design practically achieves close to the 2x theoretically expected throughput gain 4/25/2014
Why can not we use the SISO design to implement MIMO full duplex? 4/25/2014 6
State of the Art for SISO full duplex 4/25/2014 7
State of the Art for SISO full duplex X Conceptual model RX 4/25/2014 7
State of the Art for SISO full duplex Reflector X Conceptual model RX 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections X Conceptual model RX 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections R+ X Conceptual model RX 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections R+ X Conceptual model RX R + ε 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections R+ X Conceptual model RX R + ε 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections R+ Reflector R R+ X RF Frontend RX RF Frontend X Conceptual model RX R + ε X Practical Realization of RX 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections R+ Reflector RF Circuit d1 fixed delays a 1 R R+ +δ X RF Frontend d N a N tuning algorithm RX RF Frontend X Conceptual model RX R + ε X Practical Realization of RX 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections R+ Reflector RF Circuit d1 fixed delays a 1 R R+ +δ X Conceptual model RX R + ε X RF Frontend X d N a N tuning algorithm Digital Eliminates all Linear and Non-Linear Distortion Practical Realization of RX RF Frontend RX R + ε 4/25/2014 7
State of the Art for SISO full duplex Reflector Channel representing environmental reflections R+ Reflector RF Circuit d1 fixed delays a 1 R R+ +δ X Conceptual model RX R + ε X RF Frontend tuning algorithm Digital Eliminates all Linear and Non-Linear Distortion 4/25/2014 7 adaptive that closely matches the environmental X d N a N Practical Realization of RX RF Frontend RX R + ε Full duplex designs are conceptually realizing an
Key Metrics for SISO Full Duplex Design 4/25/2014 8
Key Metrics for SISO Full Duplex Design s are characterized by two key metrics 4/25/2014 8
Key Metrics for SISO Full Duplex Design s are characterized by two key metrics Complexity: he number of taps needed to closely match the environmental reflection response. igher number of taps More analog circuit area, more FPGA gates Less is better 4/25/2014 8
Key Metrics for SISO Full Duplex Design s are characterized by two key metrics Complexity: he number of taps needed to closely match the environmental reflection response. igher number of taps More analog circuit area, more FPGA gates Less is better Interference residue: Amount of interference left uncancelled, depends on the accuracy of the in modeling all reflections lower is better 4/25/2014 8
Key Metrics for SISO Full Duplex Design s are characterized by two key metrics Complexity: he number of taps needed to closely match the environmental reflection response. igher number of taps More analog circuit area, more FPGA gates Less is better Interference residue: Amount of interference left uncancelled, depends on the accuracy of the in modeling all reflections lower is better SISO Full Duplex Implications Analog taps 12 6x6 inches board Digital taps 132 295 DSP 48 Logic Interference Residue 1dB over noise floor Almost optimal full duplex 4/25/2014 8
Key Metrics for SISO Full Duplex Design s are characterized by two key metrics Complexity: he number of taps needed to closely match the environmental reflection response. igher number of taps More analog circuit area, more FPGA gates Less is better Interference residue: Amount of interference left uncancelled, depends on the accuracy of the in modeling all reflections lower is better SISO Full Duplex Implications Analog taps 12 6x6 inches board Digital taps 132 295 DSP 48 Logic Interference Residue 1dB over noise floor Almost optimal full duplex Goal for any full duplex design: Minimize interference residue to the noise floor with the lowest complexity 4/25/2014 8 cancellation
ow is the MIMO full duplex problem different? X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 12 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 12 13 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 12 13 11 12 13 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 12 13 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 12 13 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 21 22 23 11 12 13 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 21 22 23 11 13,21 12,, 23 22 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 13,21 12,, 23 22 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 13,21 12,, 23 22 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 31 32 33 11 13,21 12,, 23 22 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 31 32 33 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector A1 A2 3 Antenna Full Duplex MIMO Radio A3 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector A1 11 A2 3 Antenna Full Duplex MIMO Radio A3 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector A1 11 12 A2 A3 3 Antenna Full Duplex MIMO Radio 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? 11 12 13 Reflector A1 A2 3 Antenna Full Duplex MIMO Radio A3 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector A1 11 13 12 22 23 21 A2 A3 3 Antenna Full Duplex MIMO Radio 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector A1 11 13 31 12 22 23 21 33 A2 32 3 Antenna Full Duplex MIMO Radio A3 11 13,21, 31 12, 22,, 23, 33 32 X1 X2 X3 RX3
ow is the MIMO full duplex problem different? Reflector A1 11 13 31 12 22 23 21 33 A2 32 3 Antenna Full Duplex MIMO Radio A3 11 13,21, 31 12, 22,, 23, 33 32 X1 MIMO full duplex X2has quadratically more X3 number RX3 of signals to cancel because of the presence of cross talk.
Why not replicate the SISO full duplex design to cancel all the self-talk and cross-talk components for MIMO? 4/25/2014 10
Naïve Solution: Replicates SISO design A1 A2 3 Antenna Full Duplex MIMO Radio A3
Naïve Solution: Replicates SISO design 11 A1 A2 3 Antenna Full Duplex MIMO Radio A3
Naïve Solution: Replicates SISO design 11 A1 A2 3 Antenna Full Duplex MIMO Radio A3
Naïve Solution: Replicates SISO design 11 A1 A2 3 Antenna Full Duplex MIMO Radio A3 11 11 X1
Naïve Solution: Replicates SISO design 11 22 A1 A2 3 Antenna Full Duplex MIMO Radio A3 11 11 X1
Naïve Solution: Replicates SISO design 11 22 A1 A2 3 Antenna Full Duplex MIMO Radio A3 22 11 11 22 X1 X2
Naïve Solution: Replicates SISO design 11 22 33 A1 A2 3 Antenna Full Duplex MIMO Radio A3 22 11 11 22 X1 X2
Naïve Solution: Replicates SISO design 11 22 33 A1 A2 3 Antenna Full Duplex MIMO Radio A3 11 22 33 11 33 22 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 11 22 33 A1 21 A2 A3 3 Antenna Full Duplex MIMO Radio 11 22 33 11 33 22 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 11 22 33 A1 21 A2 A3 3 Antenna Full Duplex MIMO Radio 11 21 22 33 11 33 22 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 11 22 33 A1 21 A2 A3 3 Antenna Full Duplex MIMO Radio 11 11 21 21 22 33 22 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 22 33 A1 21 A2 A3 3 Antenna Full Duplex MIMO Radio 11 11 21 21 22 33 22 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 22 33 A1 21 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 22 33 22 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 22 33 A1 21 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 22 33 22 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 22 33 22 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 22 33 22 32 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 22 22 32 32 33 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 12 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 22 22 32 32 33 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 12 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 12 22 22 32 32 33 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 11 12 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 12 22 22 32 32 12 33 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 13 11 12 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 31 11 11 21 21 31 12 22 22 32 32 12 33 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 13 11 12 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 13 31 11 11 21 21 31 12 22 22 32 32 12 33 33 RX3 X1 X2 X3
Naïve Solution: Replicates SISO design 31 13 11 12 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 13 31 11 11 21 21 31 12 22 22 32 32 12 33 33 RX3 13 X1 X2 X3
Naïve Solution: Replicates SISO design 31 13 11 12 23 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 13 31 11 11 21 21 31 12 22 22 32 32 12 33 33 RX3 13 X1 X2 X3
Naïve Solution: Replicates SISO design 31 13 11 12 23 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 13 31 11 11 21 21 31 12 22 22 32 32 12 23 33 33 RX3 13 X1 X2 X3
Naïve Solution: Replicates SISO design 31 13 11 12 23 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 13 31 11 11 21 21 31 12 22 22 32 32 12 23 33 33 RX3 23 13 X1 X2 X3
Naïve Solution: Replicates SISO design 31 13 11 12 23 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 13 31 11 11 21 21 31 12 22 22 32 32 12 23 33 33 RX3 23 13 X1 X2 otal 9N aps, for M=3 X3
Naïve Solution: Replicates SISO design 31 13 11 12 23 33 22 21 32 A1 A2 A3 3 Antenna Full Duplex MIMO Radio 11 complexity 31 22 is quadratic with number of 33 32 21 antennas (M) 23 12 Even worse cancellation residue at each receiver 11 increases 21 linearly with number of 32 antennas (e.g. 3x3 5dB residue) 31 22 12 13 33 RX3 23 13 X1 X2 otal 9N aps, for M=3 X3
CDF Why does the increase in cancellation residue matter? hroughput performance of SISO replication design 1 0.8 Our Design 0.6 SISO Replication 0.4 0.2 SISO Low Complexity Replication 0 0 0.5 1 1.5 2 Gain relative to alf Duplex 3X3 MIMO Median throughput gain is not close to 2x because of large interference residue 4/25/2014 12
CDF Why does the increase in cancellation residue matter? hroughput performance of SISO replication design 1 0.8 Our Design 0.6 0.4 0.2 Worse than standard alf Duplex 1.35x SISO Replication SISO Low Complexity Replication 0 0 0.5 1 1.5 2 Gain relative to alf Duplex 3X3 MIMO Median throughput gain is not close to 2x because of large interference residue 4/25/2014 12
CDF Why does the increase in cancellation residue matter? hroughput performance of SISO replication design 1 0.8 0.6 0.4 0.2 Worse than standard alf Duplex Ideal Solution: 1.35x 0 0 0.5 1 1.5 2 Gain relative to alf Duplex 3X3 MIMO Our Design SISO Replication SISO Low Complexity Replication complexity scales linearly with number of antennas (M) Median throughput Residue gain is same not close as to SISO 2x because and should of not be impacted large by interference number ofresidue MIMO antennas 4/25/2014 12
echnical Contributions Designed & implemented a near-ideal MIMO in-band full duplex radio ybrid (analog & digital) cross and self-talk cancellation circuits & algorithms that completely eliminate interference to the noise floor Prototype 3x3 MIMO full duplex radios using off-the-shelf WiFi radios Experimental indoor evaluation which demonstrates that our design practically achieves close to the 2x theoretically expected throughput gain 4/25/2014
echnical Contributions Designed & implemented a near-ideal MIMO in-band full duplex radio ybrid (analog & digital) cross and self-talk cancellation circuits & algorithms that completely eliminate interference to the noise floor complexity scales linearly with M (number of antennas). Prototype 3x3 MIMO full duplex radios using off-the-shelf WiFi radios Experimental indoor evaluation which demonstrates that our design practically achieves close to the 2x theoretically expected throughput gain 4/25/2014
echnical Contributions Designed & implemented a near-ideal MIMO in-band full duplex radio ybrid (analog & digital) cross and self-talk cancellation circuits & algorithms that completely eliminate interference to the noise floor complexity scales linearly with M (number of antennas). residue is the same as the SISO full duplex, i.e. it does not degrade with increasing number of antennas. Prototype 3x3 MIMO full duplex radios using off-the-shelf WiFi radios Experimental indoor evaluation which demonstrates that our design practically achieves close to the 2x theoretically expected throughput gain 4/25/2014
Key Idea: Reducing Complexity 11 X1 11
Key Idea: Reducing Complexity 11 11 X1 X2
Key Idea: Reducing Complexity 11 21 11 X1 X2
Key Idea: Reducing Complexity 11 21 11 X1 X2 Can we reuse the self talk cancellation to also cancel the cross talk?
Why can self talk cancellation be used to partly model cross talk? 11 f 21 f 11 f X1 X2 4/25/2014 15
Why can self talk cancellation be used to partly model cross talk? Share Environment Share Reflectors 11 f 21 f 11 f X1 X2 4/25/2014 15
Why can self talk cancellation be used to partly model cross talk? Share Environment Share Reflectors 11 f 21 f Self talk and cross talk transfer function are related as they undergo similar environment. 11 f X1 X2 4/25/2014 15
Why can self talk cancellation be used to partly model cross talk? Share Environment Share Reflectors 11 f 21 f Self talk and cross talk transfer function are related as they undergo similar environment. A lot of work for modeling self-talk is already done, we need to create only the difference in self talk and cross talk. X1 11 f X2 4/25/2014 15
Why can self talk cancellation be used to partly model cross talk? Share Environment Share Reflectors Self talk and cross talk transfer function are related as they undergo similar environment. A lot of work for modeling self-talk is already done, we need to create only the difference in self talk and cross talk. 11 f 21 f 11 f X1 X2 21 f = cas cade f. 11 f 4/25/2014 15
Why can self talk cancellation be used to partly model cross talk? Share Environment Share Reflectors Self talk and cross talk transfer function are related as they undergo similar environment. A lot of work for modeling self-talk is already done, we need to create only the difference in self talk and cross talk. 11 f 21 f 11 f X1 X2 21 f = cas cade f. 11 f Can we leverage this relationship to reduce the cancellation complexity 4/25/2014 15
Empirical Observation 11 f 21 f 21 f = cas cade f. 11 f Cascade ransfer Function X1 11 f X2
Empirical Observation 11 f 21 f 21 f = cas cade f. 11 f Cascade ransfer Function X1 11 f X2 Cascade ransfer Function Collect cross talk and self talk for various indoor environments
Empirical Observation 11 f 21 f 21 f = cas cade f. 11 f Cascade ransfer Function X1 11 f X2 Cascade ransfer Function Collect cross talk and self talk for various indoor environments Learning Algorithm Complexity Reduction From all the possible cascade response, calculate via optimization the best low complexity circuit which achieves the cascade transfer function (offline analysis) hese cascade circuits are very low complexity, thus allowing us to get close to linear complexity
Reducing Complexity: Cascaded 11 21 31 11 N taps RX3 X2 X3 X1
Reducing Complexity: Cascaded 11 21 31 11 N taps RX3 c21 X2 X3 X1 C taps
Reducing Complexity: Cascaded 11 21 31 11 N taps RX3 c21 c31 X2 X3 X1 C taps D taps
Reducing Complexity: Cascaded 11 21 31 11 N taps RX3 c21 c31 X2 X3 X1 C taps D taps Cascaded s
Reducing Complexity: Cascaded 11 21 31 11 N taps RX3 c21 c31 X2 X3 X1 C taps D taps Cascaded s aps: N >> C > D
Reducing Complexity: Cascaded 11 21 31 11 N taps RX3 c21 c31 X2 X3 X1 C taps D taps Cascaded s aps: N >> C > D otal aps: N+ C + D, for Chain 1 General Complexity per chain: ~N << M.N
Complete Cascaded Design: 11 33 11 22 RX3 X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 11 21 33 11 22 RX3 X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 11 21 33 11 22 RX3 c21 C taps X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 C taps X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 C taps D taps X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 C taps D taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 C taps D taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 C taps D taps C taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 c32 C taps D taps C taps C taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 c32 C taps D taps C taps C taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 c32 c32 C taps D taps C taps C taps C taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 c32 c32 c31 C taps D taps C taps C taps C taps D taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 c32 c32 c31 C taps D taps C taps C taps C taps D taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 c32 c32 c31 C taps D taps C taps C taps C taps D taps Cascaded X1 X2 aps: N >> C > D X3
Complete Cascaded Design: 31 11 21 33 11 22 RX3 c21 c31 c12 c32 c32 c31 C taps D taps C taps C taps C taps D taps Cascaded X1 X2 X3 aps: N >> C > D otal aps: 3N+ 4C + 2D, for M=3 General Complexity: ~ M.N << M 2.N
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 A1 Analog Isolation A2 Analog Isolation A3 Analog Isolation RX3 X1 X2 xb3
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 X2 xb3
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 31 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 X2 xb3
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 31 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 C X2 xb3
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 31 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 C X2 xb3 esidue=e + 21 + 31
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 31 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 C X2 xb3 esidue=e + 21 + 31 C
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 31 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 C X2 xb3 esidue=e + 21 + 31 Residue=2e + 31 C
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 31 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 C X2 xb3 esidue=e + 21 + 31 C Residue=2e + 31 C
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design 11 31 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation RX3 X1 C X2 xb3 esidue=e + 21 + 31 C Residue=2e + 31 C Residue= 3e
nterference Residue with Cascaded Design shows no egradation Independent raining with SISO Replication Design Improved Joint raining with Cascaded Design 11 31 11 31 21 A1 Analog Isolation 21 A2 Analog Isolation A3 Analog Isolation A1 Analog Isolation A2 Analog Isolation A3 Analog Isolation RX3 X1 C X2 xb3 X1 X2 X3 RX3 esidue=e + 21 + 31 Residue=2e + 31 C C Cascaded Cancel Residue= 3e Residue = e
Evaluation Q1: Does it work with commodity radios?
Evaluation Q1: Does it work with commodity radios? Goal: Build 3X3 MIMO full duplex radio using commodity software radios.
Evaluation Q1: Does it work with commodity radios? Goal: Build 3X3 MIMO full duplex radio using commodity software radios. Maxim transceiver Challenge: Extremely high transmitter noise and nonlinearities 20Mz BW (transceiver limitation) 20 dbm max X power WiFi 802.11n PY
Evaluation Q1: Does it work with commodity radios? Goal: Build 3X3 MIMO full duplex radio using commodity software radios. 11 22 Maxim transceiver Challenge: Extremely high transmitter noise and nonlinearities 20Mz BW (transceiver limitation) 20 dbm max X power WiFi 802.11n PY c12 c21
Evaluation Q1: Does it work with commodity radios? 4/25/2014 21
Evaluation Q1: Does it work with commodity radios? Indoor noisy office environment 4/25/2014 21
Power in dbm Evaluation Q1: Does it work with commodity radios? Indoor noisy office environment 20 0-20 -40-60 -80 2.43 2.44 2.45 2.46 2.47 Frequency in Ghz otal X Signal After Self-talk Cancel After Cross-talk 1 Cancel After Cross-talk 2 Cancel After Digital Cancel Noise floor 4/25/2014 21
Power in dbm Evaluation Q1: Does it work with commodity radios? Indoor noisy office environment unes to environmental changes within 8us, needs to be re-tuned every 60 ms 20 0-20 -40-60 -80 2.43 2.44 2.45 2.46 2.47 otal X Signal Frequency in Ghz Prototype implementation completely cancels interference to noise floor for a 3x3 MIMO radio 4/25/2014 21 After Self-talk Cancel After Cross-talk 1 Cancel After Cross-talk 2 Cancel After Digital Cancel Noise floor
Evaluation Q2: Does it have linear Complexity Scaling? Resource Comparison between SISO replication and our design Analog taps (3X3) Digital taps (3X3) SISO replication design Our design 108 56 (reduced by 1.92x) 1188 485 (reduced by 2.45x) uning time (3X3) 9 ms.024 ms (reduced by 375x)
Evaluation Q2: Does it have linear Complexity Scaling? Resource Comparison between SISO replication and our design Analog taps (3X3) Digital taps (3X3) SISO replication design Our design 108 56 (reduced by 1.92x) 1188 485 (reduced by 2.45x) uning time (3X3) 9 ms.024 ms (reduced by 375x) Complexity of both Analog and Digital, scales linearly as number of antennas increases
CDF Evaluation Q4: Does that translate to doubling of throughput in practice? 1 0.8 0.6 Our Design SISO Replication 0.4 0.2 SISO Low Complexity Replication 0 0 0.5 1 1.5 2 Gain relative to alf Duplex 3X3 MIMO 4/25/2014 23
CDF Evaluation Q4: Does that translate to doubling of throughput in practice? 1 0.8 0.6 Our Design SISO Replication 0.4 0.2 1.95x SISO Low Complexity Replication 0 0 0.5 1 1.5 2 Gain relative to alf Duplex 3X3 MIMO 4/25/2014 23
CDF Evaluation Q4: Does that translate to doubling of throughput in practice? 1 0.8 Our Design 0.6 0.4 0.2 Worse than standard alf Duplex 1.95x SISO Replication SISO Low Complexity Replication 0 0 0.5 1 1.5 2 Gain relative to alf Duplex 3X3 MIMO Our design practically achieves the theoretical throughput doubling 4/25/2014 23
Conclusion Design and implementation of a near-ideal complexity and performance full duplex MIMO radio Shows that full duplex and MIMO can operate concurrently as applications to many other problems Radio slicing, backscatter, imaging etc Many implications for MAC layer design Feedback, beamforming, MU-MIMO 4/25/2014 24