1. High-Throughput Enhancements for 802.11: PHY Supplement
Month 2002
doc.: IEEE /xxxr0
November 2004
High-Throughput Enhancements for : PHY Supplement
John Ketchum, Sanjiv Nanda, Rod Walton, Steve Howard, Mark Wallace
QUALCOMM, Incorporated 9 Damonmill Square, Suite 2A Concord, MA Phone: Fax:
John Ketchum et al, Qualcomm
John Doe, His Company

2. Agenda Calibration Eigenvalue decomposition complexity
November 2004
Agenda
Calibration
Eigenvalue decomposition complexity
Eigenvector steering modes of operation
PHY Q&A
John Ketchum et al, Qualcomm

3. Over-the-Air Calibration
November 2004
Over-the-Air Calibration
ES approach requires over-the-air calibration procedure
Compensates for amplitude and phase differences in Tx and Rx chains
Calibration required infrequently– typically on association only
Simple exchange of calibration symbols and measurement information requires little overhead and background processing
Total of ~1000 bytes of calibration data exchanged for 2x2 link
~2800 bytes for 4x4 link
John Ketchum et al, Qualcomm

4. Over-the-Air Calibration Procedure
November 2004
Over-the-Air Calibration Procedure
Calibration typically is initiated by client STA on association or when STA determines that it needs calibration due to performance degradation
AP uses repeated calibrations with many clients to refine its calibration
Both STAs involved in calibration procedure compute channel estimate
Only one STA computes calibration vectors—keeps one set and sends the other to the other STA
John Ketchum et al, Qualcomm

5. Measured Calibration Errors
Month 2002
doc.: IEEE /xxxr0
November 2004
Measured Calibration Errors
Calibration based on a single channel measurement in each directions
Measured calibration error is a strong function of SNR
Resulting calibration used in operating modem
Residual errors are compensated by receive processing
John Ketchum et al, Qualcomm
John Doe, His Company

6. Achieved Average Calibration Error
November 2004
Achieved Average Calibration Error
Shows average of repeated calibration procedures
Demonstrates that calibration error is primarily limited by receiver noise
John Ketchum et al, Qualcomm

7. Degradation Due to Calibration Error
November 2004
Degradation Due to Calibration Error
Shows cumulative distribution function of capacity in 4×4 on a random flat fading channel
Average receive SNR is 20 dB; average calibration error is -10 dB
Less than 10% degradation in median capacity
John Ketchum et al, Qualcomm

8. Degradation Due to Calibration Error
November 2004
Degradation Due to Calibration Error
Shows cumulative distribution function of capacity in 4×4 on a random flat fading channel
Average receive SNR is 20 dB; average calibration error is -20 dB
Minimal degradation in median capacity
John Ketchum et al, Qualcomm

9. Degradation Due to Calibration Error
November 2004
Degradation Due to Calibration Error
Shows cumulative distribution function of capacity in 4×4 on a random flat fading channel
Average receive SNR is 10 dB; average calibration error is -10 dB
Minimal degradation in median capacity
John Ketchum et al, Qualcomm

10. TDD Reciprocal Channel
November 2004
TDD Reciprocal Channel
In a TDD MIMO system, the over-the-air portion of the channel is reciprocal
The up-link channel, , (entity A to entity B) is the transpose of the down-link channel, , ( is the OFDM sub-carrier index):
Due to gain differences in Tx and Rx chains at both ends of the link, the baseband-to-baseband channel is not reciprocal.
The observed channel is weighted by two diagonal matrices with the complex gains of the transmit and receive chains:
These gain differences can be removed with a simple over-the-air calibration procedure that learns the gain matrices
Result is a very stable calibrated reciprocal channel
John Ketchum et al, Qualcomm

11. Calibration Procedure
November 2004
Calibration Procedure
Find diagonal calibration matrices that can be applied to transmit vectors to compensate for amplitude/phase variations between Rx and Tx chains in device
Calibration required once per session; i.e., upon association
Procedure as follows:
Entity A (typically a STA) observes MIMO pilot from entity B (typically an AP)
Entity A forms an estimate of channel,
Entity A transmits MIMO pilot, which entity B observes
Entity B forms channel estimate,
Entity B transmits to A
Requires transmitting bit values
624 B for 2x2; B for 4x4
John Ketchum et al, Qualcomm

12. Calibration Procedure
November 2004
Calibration Procedure
Entity A now has both
Entity A now solves for diagonal calibration matrices such that
Entity A sends to entity B, then both ends of link have calibration matrix
Requires transmitting bit values
624 B for 4 antennas; 312 B for 2 antennas
Calibration matrices are incorporated into Tx steering vectors.
John Ketchum et al, Qualcomm

13. Calibration Summary Calibration required infrequently
November 2004
Calibration Summary
Calibration required infrequently
Low overhead message exchange
Non-time-critical channel estimates and calibration vector calculations can be performed as background tasks
Calibration errors have minimal impact on performance
Receiver processing required to clean up other misalignments takes care of calibration errors as well
John Ketchum et al, Qualcomm

14. Eigenvalue Decomposition Complexity
November 2004
Eigenvalue Decomposition Complexity
Only one device in a corresponding pair needs to calculate eigenvalue decomposition (EVD)
Full singular value decomposition (SVD) is not required—only right singular vectors for transmit steering
Can be computed as eigenvectors of channel correlation matrix
John Ketchum et al, Qualcomm

15. Eigenvalue Decomposition Complexity
November 2004
Eigenvalue Decomposition Complexity
Very low complexity direct calculation for two-antenna device
Calculate eigenvalues & eigenvectors of 2×2 Hermitian matrix for each subcarrier
21 real multiplies, 3 inverts, 3 square roots required for each subcarrier
Total of 1176 multiplies, 168 inverts, 168 square roots for 56 subcarrier system
20-30 µs with minimal hardware resources
John Ketchum et al, Qualcomm

16. Eigenvalue Decomposition Complexity
November 2004
Eigenvalue Decomposition Complexity
Larger matrices (more than two antennas) require iterative calculation of EVD
Accuracy requirements are minimal compared to many other applications
Current hardware prototype completes 4×4 EVD for 52 subbands in ~800 µs on FPGA running at 80 MHz with modest hardware resources.
John Ketchum et al, Qualcomm

17. Wideband Eigenmodes on TDD Reciprocal Channel
November 2004
Wideband Eigenmodes on TDD Reciprocal Channel
Uplink channel SVD:
Tx steering matrix:
Rx matched filter:
Downlink SVD:
: downlink Tx steering matrix
Transmit steering vectors at one end of the link can be computed directly from the receive matched filter at the same end of the link
Normalize and conjugate
Since Tx steering vectors can be obtained directly from Rx matched filter, eigenvectors only need to be computed at one end of the link
John Ketchum et al, Qualcomm

18. Distributed Computation of SVDs
November 2004
Distributed Computation of SVDs
Client STAs compute SVD
Relieves AP of burden of SVD computation for many STAs
Under EDCA access rules, STAs send a training request (TRQ) to AP
AP responds with PPDU containing unsteered reference (TRSP)
STA can use any TRSP to initiate SVD calculation
STA with data to send listens for a TRSP, and if none in reasonable period, sends TRQ
AP with data to send can transmit unsolicited TRSP addressed to data destination
Busy AP can send frequent broadcast TRSPs (~2–4 ms intervals)
STA receiving TRSP calculates SVD, and when complete, sends data in ES mod
STA with backlogged data for several PPDUs can include TRQ with data PPDU
John Ketchum et al, Qualcomm

19. ES with Distributed SVD Calculation under EDCA
November 2004
ES with Distributed SVD Calculation under EDCA
John Ketchum et al, Qualcomm

20. Distributed Computation of SVDs
November 2004
Distributed Computation of SVDs
Under ACF, TRSP is part of SCHED message
All client STAs can share the unsteered reference in the PLCP header to calculate a channel estimate and then SVD
Protocol overhead for supporting ES mode is minimized in this manner
John Ketchum et al, Qualcomm

21. ES Mode in Scheduled Access Period
November 2004
ES Mode in Scheduled Access Period
John Ketchum et al, Qualcomm

22. AP-Centric ES Mode—single TXOP
November 2004
AP-Centric ES Mode—single TXOP
AP does all SVD calculations
Substantial processing burden at AP
Tight time constraints for completion of SVD
Requires additional overhead of RTS/TRQ & CTS/TRSP
John Ketchum et al, Qualcomm

23. AP-Centric ES Mode—Relaxed Timing
November 2004
AP-Centric ES Mode—Relaxed Timing
AP “primes the pump” with TRQ, STA responds with TRSP
Result is relaxed timing compared to previous
Still requires extra overhead
John Ketchum et al, Qualcomm

24. Eigenvector Steering Operation Summary
November 2004
Eigenvector Steering Operation Summary
Eigenvector steering operation can be supported within existing EDCA and HCCA access mechanisms
Lowest protocol overhead incurred within ACF construct
Protocol overhead due to MIMO training for SVD is minimal when AP sends common training PPDU, for SVD calculation at client
Distributing SVD calculation to clients also relieves intense processing burden at AP
AP-centric ES operation, including SVD calculation, can be supported with relaxed timing constraints.
John Ketchum et al, Qualcomm

25. Interdigital Questions
November 2004
Interdigital Questions
Have you looked at some additional coding on top of the Eigen-Beamforming?
There is no need for further coding (other than either legacy BCC or advanced coding such as turbo or ldpc) on top of ES mode.
ES is optimal in the sense that it is provides optimal Tx and Rx spatial processing for a fully informed transmitter. In that sense, further tx processing such as Space-Time Block Coding or Space Frequency block coding would not be productive.
What constraints are imposed on the MAC in order to use CSI feedback?
CSI feedback can be productively utilized within the constraints of EDCA, HCCA, or ACF.
CSI feedback does not constrain the MAC. CSI feedback is available as part of the MIMO training in the PLCP header, which is common to all proposals.
Why is throughput for 4x4 more than twice that of 2x2?
In general, in the simulation results that we report, throughput for 4x4 is less than or equal to twice the throughput for 2x2. There are a few cases where this does not hold, which is due to suboptimal coding and rate selection.
John Ketchum et al, Qualcomm