PentaG™ AI Processor for CSI Reporting in 5G NR White Paper

CEVA®

PentaG™ AI Processor for CSI Reporting in 5G NR White Paper

Rev. 1.0 February 2018

CEVA® PentaG™ AI Processor for CSI Reporting in 5G NR White Paper

Documentation Control History Table

Version Date Description Remarks

1.0 26 February 2018 Initial release

Rev. 1.0 Copyright © 2018 – CEVA®, Inc. i


Disclaimer and Proprietary Information Notice The information contained in this document is subject to change without notice and does not represent a commitment on any part of CEVA®, Inc. CEVA®, Inc. and its subsidiaries make no warranty of any kind with regard to this material, including, but not limited to implied warranties of merchantability and fitness for a particular purpose whether arising out of law, custom, conduct or otherwise.

While the information contained herein is assumed to be accurate, CEVA®, Inc. assumes no responsibility for any errors or omissions contained herein, and assumes no liability for special, direct, indirect or consequential damage, losses, costs, charges, claims, demands, fees or expenses, of any nature or kind, which are incurred in connection with the furnishing, performance or use of this material.

This document contains proprietary information, which is protected by U.S. and international copyright laws. All rights reserved. No part of this document may be reproduced, photocopied, or translated into another language without the prior written consent of CEVA®, Inc.

CEVA®, CEVA-XC™, CEVA-XC5™, CEVA-XC321™, CEVA-XC323™, CEVA-Xtend™, CEVA-XC4000™, CEVA-XC4100™, CEVA-XC4200™, CEVA-XC4210™, CEVA-XC4400™, CEVA-XC4410™, CEVA-XC4500™, CEVA-XC12™, CEVA-TeakLite™, CEVA-TeakLite-II™, CEVA-TeakLite-III™, CEVA-TL3210™, CEVA-TL3211™, CEVA-TeakLite-4™, CEVA-TL410™, CEVA-TL411™, CEVA-TL420™, CEVA-TL421™, CEVA-Quark™, CEVA-Teak™, CEVA-X™, CEVA-X1620™, CEVA-X1622™, CEVA-X1641™, CEVA-X1643™, Xpert-TeakLite-II™, Xpert-Teak™, CEVA-XS1100A™, CEVA-XS1200™, CEVA-XS1200A™, CEVA-TLS100™, Mobile-Media™, CEVA-MM1000™, CEVA-MM2000™, CEVA-SP™, CEVA-VP™, CEVA-MM3000™, CEVA-MM3100™, CEVA-MM3101™, CEVA-XM™, CEVA-XM4™, CEVA-XM6™, CEVA-X1™, CEVA-X2™, CEVA-NeuPro™, CEVA-Audio™, CEVA-HD-Audio™, CEVA-VoP™, CEVA-Bluetooth™, CEVA-SATA™, CEVA-SAS™, CEVA-Toolbox™, SmartNcode™, PentaG™, and ClearVox™ are trademarks of CEVA, Inc.

All other product names are trademarks or registered trademarks of their respective owners.

ii Copyright © 2018 – CEVA®, Inc. Rev. 1.0


Support CEVA® makes great efforts to provide a user-friendly software and hardware development environment. Along with this, CEVA provides comprehensive documentation, enabling users to learn and develop applications on their own. Due to the complexities involved in the development of DSP applications that might be beyond the scope of the documentation, an online Technical Support Service has been established. This service includes useful tips and provides fast and efficient help, assisting users to quickly resolve development problems.

How to Get Technical Support:

● FAQs: Visit our website http://www.ceva-dsp.com or your company's protected page on the CEVA website for the latest answers to frequently asked questions.

● Application Notes: Visit our website http://www.ceva-dsp.com or your company's protected page on the CEVA website for the latest application notes.

● Email: Use the CEVA central support email address [email protected]. Your email will be forwarded automatically to the relevant support engineers and tools developers who will provide you with the most professional support to help you resolve any problem.

● License Keys: Refer any license key requests or problems to [email protected]. For SDT license keys installation information, see the SDT Installation and Licensing Scheme Guide.

Email: [email protected]

Visit us at: www.ceva-dsp.com

Rev. 1.0 Copyright © 2018 – CEVA®, Inc. iii

http://www.ceva-dsp.com/


mailto:[email protected]






List of Sales and Support Centers Israel USA Ireland Sweden

2 Maskit Street P.O. Box 4047 Herzelia 46120 Israel

1174 Castro Street Suite 210 Mountain View, CA 94040 USA

18/19 South William Street 2nd Floor Dublin 2 Ireland

Klarabergsviadukten 70 Box 70396 107 24 Stockholm Sweden

Tel: +972 9 961 3700 Fax: +972 9 961 3800

Tel: +1-650-417-7923 Fax: +1-650-417-7924

Tel: +353 1 237 3900 Fax: +353 1 237 3923

Tel: +46(0)8 506 362 24 Fax: +46(0)8 506 362 20

China (Shanghai)

China (Beijing)

China (Shenzhen)

Hong Kong

Unit 1203, Building E Chamtime Plaza Office Lane 2889, Jinke Road Pudong New District Shanghai, 201203 China

Rm 503, Tower C Raycom InfoTech Park No. 2, Kexueyuan South Road Haidian District Beijing 100190 China

Rm 709, Tower A SCC Financial Centre No. 88 First Haide Avenue Nanshan District Shenzhen 518064 China

Level 43, AIA Tower 183 Electric Road North Point Hong Kong

Tel: +86-21-20577000 Fax: +86-21-20577111

Tel: +86-10 5982 2285 Fax: +86-10 5982 2284

Tel: +86-755-84356027 Tel: +852-39751264

South Korea Taiwan Japan France

#478, Hyundai Arion 147, Gumgok-Dong Bundang-Gu Sungnam-Si Kyunggi-Do, 463-853 South Korea

Room 621 No. 1, Industry E, 2nd Rd Hsinchu, Science Park Hsinchu 300 Taiwan R.O.C

1-6-5 Shibuya SK Aoyama Bldg. 3F Shibuya-ku, Tokyo 150-0002 Japan

RivieraWaves S.A.S 400, avenue Roumanille Les Bureaux Green Side 5, Bât 6 06410 Biot - Sophia Antipolis France

Tel: +82-31-704-4471 Fax:+82-31-704-4479

Tel: +886 3 5798750 Fax: +886 3 5798750

Tel: +81-3-5774-8250 Tel: +33 4 83 76 06 00 Fax: +33 4 83 76 06 01

iv Copyright © 2018 – CEVA®, Inc. Rev. 1.0


Table of Contents

1. INTRODUCTION .................................................................................................. 1 1.1 Challenges toward 5G NR Release 15 .................................................................1 1.2 Beamforming Schemes .....................................................................................2 1.3 PentaG Solution and the AI Processor .................................................................4

2. LINK PERFORMANCE ABSTRACTION .................................................................... 5 2.1 Link Performance Abstraction by MIB Metric ........................................................7

3. CEVA AI PROCESSOR FOR 5G NR CSI REPORTING ............................................... 9 3.1 Performance Simulation ....................................................................................9

4. REFERENCES ..................................................................................................... 13

5. GLOSSARY ......................................................................................................... 15

List of Examples Example 2-1: Cross-Sections of MIB for MIMO Rank 2 Channel ...........................................6 Example 2-2: Link Performance Abstraction Procedure by MIB Metric ..................................7

List of Figures Figure 1-1: 5G NR Beamforming Schemes ........................................................................2 Figure 1-2: PentaG Hardware Architecture .......................................................................4 Figure 3-1: Performance Results (Medium Correlation) .................................................... 10 Figure 3-2: Performance Results (High Correlation) ......................................................... 10

List of Tables Table 5-1: Acronyms ................................................................................................... 15

Rev. 1.0 Copyright © 2018 – CEVA®, Inc. v


1. Introduction 1.1 Challenges toward 5G NR Release 15

Compared to previous LTE releases, 5G NR introduces a significant increase in the throughput, latency, and number of users, as well as in the connection density.

Because 5G NR is targeted at supporting the entire frequency band, from low band in UHF up to 100 GHz, a number of frequency-related signal-propagation challenges must be addressed:

● Path loss increases with frequency (because the wavelength decreases with frequency).

● Signal diffraction loss increases with frequency. ● Air interface and atmosphere losses are frequency dependent. ● Scattering signal increases with frequency. ● Penetration loss increases with frequency

The channel model presented in 3GPP TR38.901 is based on 3D channel characteristics, and includes additional features like antenna model and more channel scenarios. All of these features, in addition to these frequency-related challenges, influence the channel model that must be used with 5G NR transmissions. This also means that the CSI must support wide flexibility in its link adaptation functionalities.

One of the key requirements in 5G NR is support for massive MIMO technology as an answer to the wide frequency range, as well as the opportunity to exploit the channel characteristics and the extremely wide frequency band. In addition, the massive MIMO technology enables working in MU-MIMO and increases the overall system throughput.

Although massive MIMO is supported in Release 14 with 32 ports, the 5G NR massive MIMO started in Release 15 with 32 ports, and is expected to increase up to 64 antennas (or more) in future releases.

Rev. 1.0 Copyright © 2018 – CEVA®, Inc. 1


1.2 Beamforming Schemes To exploit the massive MIMO antennas, one of the main enhancements in 5G NR transmissions are the beamforming schemes. Depending on the antenna-array architecture (as shown in Figure 1-1), the following beamforming schemes are supported:

● Analog ● Hybrid ● Digital

Figure 1-1: 5G NR Beamforming Schemes

The different beamforming schemes require advanced CSI mechanisms to handle the specific architecture. The MU-MIMO approaches enabled in 5G NR allow up to 12 co-scheduling users to work in the same frequency range and at the same time.

In 5G NR, in addition to the transmission architecture, the channels' state information requires the use of a codebook to work with multi-panel antennas.

gNodeB requests the following indications from the UE, which sends them back as CSI reports:

● Channel Quality Indication (CQI) ● Precoding Matrix Indication (PMI) ● CSI-RS Resource Indication (CRI) ● Strongest Layer Indication (SLI) ● Rank Indication (RI) ● L1-RSRP

2 Copyright © 2018 – CEVA®, Inc. Rev. 1.0


To handle reports in such complex scenarios while supporting multiple antenna characteristics, 5G NR beam management and the 5G NR CSI procedures must support the following main codebook types:

● Type I: SU-MIMO CSI, including the following: ○ Single-panel codebooks:

CSI reports based on Non-Precoded CSI-RS ■ Supports single-user and multi-user MIMO ■ Contains 1-8 MIMO layers ■ Supports wide- and sub-band precoder selection and ■

co-phasing ○ Multi-panel codebooks:

CSI reports based on Non-Precoded CSI-RS ■ Based on single CSI method with inter-panel co-phasing ■

support Supports up to four antenna panels ■

● Type II: MU-MIMO CSI, including the following: ○ Non-Precoded CSI:

Linear combination for different DFT beams based on a ■predefined codebook

Contains 1-2 MIMO layers ■ Supports wide- and sub-band precoder selection and ■

co-phasing ○ Precoded CSI:

UE selects a set of ports when each port is beamformed using ■a linear combination of selected beams

Contains 1-2 MIMO layers ■ Supports wide- and sub-band precoder selection and ■

co-phasing

In summary, the link adaptation procedures for 5G NR must support a wide range of complex CSI and beam-management processes. In addition, the maximal latency for processing CSI reports in the UE was decreased dramatically (as compared to previous LTE releases) due the shortened 5G NR slot.



1.3 PentaG Solution and the AI Processor The PentaG™ IP platform was created to cope with the challenges introduced by the new 5G NR technology, in particular, the CSI reporting challenges described in Section 1.1.

PentaG is the industry's first 5G NR IP platform capable of meeting the extreme performance, low latency, and strict power budget requirements of the 3GPP 5G NR UE solution for eMBB devices. The PentaG architecture contains specialized scalar and vector DSP processors, co-processors, an AI processor, accelerators, software, and other essential IP blocks, in a highly configurable and modular architecture.

Figure 1-2 shows a high-level architecture of the PentaG IP platform and the AI processor that is a part of it. For more details, about the PentaG, see Ref. #[ 7].

Figure 1-2: PentaG Hardware Architecture

The AI processor that is one of the components of the PentaG solution is a fully programmable and scalable processor that uses machine-learning methods to enable advanced beamforming techniques and link adaptation scenarios (as defined in 5G NR Release 15, Type-I and Type-II) via a neural network (NN). The AI processor increases downlink data throughput and ensures unmatched performance under the most advanced maximum likelihood decoding (MLD) equalization scheme



2. Link Performance Abstraction The procedure of downlink CSI reporting in 5G NR requires a report of the preferred configuration for MIMO transmission. The UE must select the optimal MIMO configuration parameters, for example, PMI, CQI, and so on.

To choose the best configuration among a large set of all possible combinations, the UE must convert the MIMO channel conditions (which are typically a function of several parameters) into a single scalar metric. This metric can be used afterwards to compare different MIMO parameter configurations. For example, a channel metric can be compared against predefined thresholds to predict whether the Block Error Rate (BLER) is expected to exceed 10%.

The following link performance abstraction metrics are the most common:

● Effective Exponential SNR Mapping (EESM): Estimates the effective SNR. This metric is suitable for linear receivers where receiver performance is dictated by the effective SNR, and MIMO inter-layer interferences are regarded as equivalent noise.

● Mutual Information per Bit (MIB): Estimates the average amount of information that the Log Likelihood Radio (LLR) at the receiver contains about transmitted bits, as follows:

𝑴𝑴𝑴𝑴𝑴𝑴(𝒃𝒃;𝑳𝑳𝑳𝑳𝑳𝑳) = � 𝑴𝑴(𝒃𝒃𝒊𝒊;𝑳𝑳𝑳𝑳𝑳𝑳𝒊𝒊) 𝑴𝑴

𝒊𝒊=𝟏𝟏

Where: ○ M is the number of bits in a QAM constellation. ○ MIB values are in range [0,1].

In contrast to the EESM, the MIB metric does not have an exact analytic formula. However, several approximations have been proposed:

● In Ref. #[ 5], an exact formula for MIB was derived for an AWGN (with no MIMO interference) channel and a BPSK/QPSK-modulated signal. For higher QAM constellations and AWGN channels, an approximate formula was proposed.

● In Ref. #[ 6], an approximated formula was proposed for calculating the MIB of an MIMO channel with two layers. This formula relies on knowledge of the eigenvalues and the condition number of the MIMO channel matrix. However, for high correlation MIMO channels, this approximation might not be accurate.



Example 2-1 demonstrates cross-sections of an MIB for an MIMO rank 2 channel.

Example 2-1: Cross-Sections of MIB for MIMO Rank 2 Channel



2.1 Link Performance Abstraction by MIB Metric Example 2-2 demonstrates a link performance abstraction procedure according to an MIB metric.

Example 2-2: Link Performance Abstraction Procedure by MIB Metric

In this example, the following occurs:

The MIB metric for each subcarrier is computed from the knowledge of 1.the MIMO channel in UE.

All of the MIB values per frequency band are averaged using (simple) 2.arithmetic averaging to get the mean MIB (MMIB).

The MMIB metric is compared against predefined thresholds to 3.determine if the BLER on that channel is expected to exceed 10%.

The combination of per-codeword CQIs that achieves maximal 4.efficiency with BLER < 10% is selected.

This procedure is repeated for each frequency band and for each hypothesis of a MIMO precoder and rank.



3. CEVA AI Processor for 5G NR CSI Reporting Neural Networks (NNs) are widely used in many engineering disciplines, for example, object recognition in images, voice recognition, natural language processing, and others. The tremendous success of an NN stems mainly from its ability to build efficient computation structures that can be trained to perform classification and/or function regression tasks. An NN can be used to solve optimization problems by observing the examples when there is no analytic formula that describes the optimal solution, or even when there is no defined optimality criteria. In the training stage, the NN learns about the hidden patterns of the data from examples.

An NN can be trained to provide any type of CSI report, as described in Section 1. The CEVA AI processor for CSI reporting in 5G NR uses an NN, which provides the following advantages:

● Flexibility: During the training stage, the NN learns to predict optimal MIMO configuration parameters for any type of link abstraction metric. A dataset for NN training is generated by PHY link layer Monte-Carlo simulations, which enables the NN to be adjusted for any type of receiver.

● Scalability: The accuracy of prediction can be traded for the size of an NN, and a more complex NN will provide better performance. Depending on the scenario, fully or partially connected NN architectures can be used. The most suitable type of NN is automatically selected at runtime for optimization of power consumption (patent pending).

The CEVA AI processor NN can be trained using any standard training toolbox (for example, the MATLAB NN toolbox or Google's TensorFlow).

3.1 Performance Simulation Figure 3-1 and Figure 3-2 show a downlink simulation comparison between the average achievable throughput using a CSI procedure via the CEVA AI processor, and the achievable throughput using an EESM metric.

Note: The EESM metric is computed without the use of the CEVA AI processor.

In both cases, a near-ML decoder was used for a MIMO rank 4 channel. Throughput was measured for medium and high correlation channels (as defined in 36.101).



Figure 3-1: Performance Results (Medium Correlation)

Figure 3-2: Performance Results (High Correlation)



The difference in performance between the two CSI reporting algorithms is more prominent at higher SNR levels and at higher correlation channels. A CSI that uses the CEVA AI processor with the MIB metric achieves significant gain in throughput relative to the EESM metric.



4. References 3GPP TS 38.211 V15.0.0: "NR; Physical channels and modulation". 1.

3GPP TS 38.212 V15.0.0: "NR; Multiplexing and channel coding". 2.

3GPP TS 38.213 V15.0.0: "NR; Physical layer procedures for control”. 3.

3GPP TS 38.214 V15.0.0: "NR; Physical layer procedures for data". 4.

IEEE 802.16m-08/004r1 clause 4, Link to System Mapping, 5.http://ieee802.org/16/tgm/docs/80216m-08_004r2.pdf

Krishna Sayana, Jeff Zhuang, and Ken Stewart, "Short Term Link 6.Performance Modeling for ML Receivers with Mutual Information per Bit Metrics", IEEE GLOBECOM 2008

CEVA PentaG™ 5G NR IP UE Platform: 7.https://www.ceva-dsp.com/product/ceva-pentag/


http://ieee802.org/16/tgm/docs/80216m-08_004r2.pdf

https://www.ceva-dsp.com/product/ceva-pentag/


5. Glossary Table 5-1 defines the acronyms used in this document.

Table 5-1: Acronyms

Term Definition

AWGN Additive White Gaussian Noise

BLER Block Error Rate

BPSK Binary Phase Shift Keying

CQI Channel Quality Indication

CRI CSI-RS Resource Indication

CSI Channel State Information

EESM Effective Exponential SNR Mapping

LLR Log Likelihood Ratio

MIB Mutual Information per Bit

MIMO Multiple Input Multiple Output

ML Maximum Likelihood

MLD Maximum Likelihood Decoding

MMIB Mean Mutual Information per Bit

NN Neural Network

PMI Precoding Matrix Indication

QAM Quadrature Amplitude Modulation

QPSK Quadrature Pulse Shift Keying

RI Rank Indication

RS Reference Signal

RSRP Reference Signal Received Power

SLI Strongest Layer Indication

SNR Signal-to-Noise Ratio

UE User Equipment


PentaG™ AI Processor for CSI Reporting in 5G NR White Paper

Documents

Transcript of PentaG™ AI Processor for CSI Reporting in 5G NR White Paper