Introduction
The performance specification for 5G is much more demanding than for previous generations of mobile networks, requiring high data rates, (up to 20 Gbps), very low latency, (< 1 mSec), ultra-high reliability, energy efficiency and very high numbers of connected devices. To meet these challenges a completely new air interface has been developed for 5G, 5G New Radio, (5G NR) which must support frequencies from low band all the way up to 100 GHz.
MU-MIMO and beamforming are two core techniques used within 5G networks which lead to multiple signal paths between base station and user equipment, each with different propagation characteristics. Channel state information (CSI) is collected for each potential signal path and then used by the Link Performance Adaptation Process to determine the optimum signal path for the specific UE.
This article examines these techniques, describes how neural networks are well suited to the computational requirements of Link Performance Adaptation and then describes how CEVA’s PentaG™ IP platform can address the CSI reporting and other challenges presented by 5G NR.
MU MIMO and beamforming
MU MIMO employed by 5G NR in order to increase throughput, uses a large number of antennas in the base station array, enabling multiple, spatially separated users to use the same time and frequency resources. 3GPP Release 15 defines 32 antennas but future releases of the 3GPP specification are expected to require 64 or more.
Beamforming, used in combination with MU MIMO, enables the radiation pattern of the antenna array to be adapted to a particular scenario, steering a lobe of power in a particular direction toward a user. Three different beamforming schemes are supported by 5G NR, depending on the architecture of the antenna-array, as shown in figure 1.
Figure 1: 5G NR Beamforming Schemes
Data transmitted between a MIMO antenna and user equipment (UE) is affected by the surrounding environment, particularly at high transmission frequencies. For example, buildings and other obstacles may cause reflections of the original signal, resulting in multiple transmission paths, each with its own signal characteristics, such as delay, attenuation, and direction of arrival.
CSI and link performance abstraction
Channel state information (CSI) is used to define the known channel properties of a communication link, describing how a signal propagates from transmitter to receiver. and represents the combined effect of, for example, scattering, fading, and power decay with distance. Effectively a collection of the spatial transfer functions between each antenna and each user terminal, CSI information is gathered in a matrix (H), as shown in Figure 2.
Figure 2: Channel State Information used to Characterise a massive MIMO system
Downlink CSI reporting is used in 5G NR to enable the UE to select the best transmission path from all of the possible options, as represented by the various configuration parameters in the CSI matrix.
To do this in a computationally efficient way, the Link Performance Abstraction process converts the MIMO channel conditions, into a single, scalar metric which can then be used to predict the Block Error Rate, (BLER), of a given channel.
Effective Exponential SINR metric is a well-known approach for link abstraction which is suitable for linear receivers but has disadvantages when used in more complex array systems. A great deal of research has been conducted into link performance methods and Mutual Information per Bit (MIB) has emerged as a process which enables a higher computational accuracy with reduced overhead.
Figure 3 illustrates the 2-stage MIB process; in the first, compression stage, a large number of quality‐related measures is reduced to a single scalar value which is then input to stage 2 where mapping is used estimate the BLER.
Figure 3: Link Performance Abstraction Procedure by MIB Metric
CEVA AI processor for 5G NR CSI reporting
Neural networks (NNs) are widely used in many engineering disciplines and their success is based on their ability to solve problems for which there is no analytic formula by learning about hidden patterns in data sets.
NNs are therefore ideally suited for CSI reporting techniques, such as MIB, and a number of NN – based processing platforms, such as CEVA’s PentaG™ IP device, (figure 4), are emerging to address the CSI reporting and other challenges presented by 5G NR.
The modular architecture of the PentaG platform contains specialized scalar and vector DSP processors, co-processors, an AI processor, accelerators, software, and other essential IP blocks. The device has been designed to meet the extreme performance, low latency, and strict power budget requirements of the 3GPP 5G NR UE solution for eMBB devices.
Figure 4: PentaG Hardware Architecture
The CEVA AI processor at the heart of the PentaG IP platform uses machine-learning methods to support advanced beamforming techniques by performing link adaptation computations using a neural network (NN).
The CEVA AI processor is adapted for 5G NR CSI reporting and offers the following advantages:
- Flexibility: During the training stage, the NN learns to predict optimal MIMO configuration parameters for any type of link abstraction metric.
The NN can be adjusted for any type of receiver, using a dataset for NN training generated by PHY link layer Monte-Carlo simulations.
- 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).
A number of standard training toolboxes, such as the MATLAB NN toolbox or Google’s TensorFlow can be used to train the CEVA AI processor.
Link adaptation in 5G NR involves a number of complex CSI computational processes whilst respecting the dramatically reduced latency performance requirements of the 5G specification. Computational throughput is therefore critical and figures 5 and 6 show a comparison between the average achievable downlink throughput using a CEVA AI processor to calculate the MIB metric, and the achievable throughput using an EESM metric, computed without the use of the CEVA AI processor. (The throughput was measured for medium and high correlation channels)
Conclusion
Massive MIMO and beamforming techniques deployed in 5G NR lead to complex CSI reporting requirements requiring sophisticated computational methods. The learning capabilities of neural networks are well suited to the requirements of CSI reporting and CEVA’s PentaG IP is an example of a technology that has been developed specifically to support 5G UE devices.
Figure 5: Performance Results (Medium Correlation)
Figure 6: Performance Results (High Correlation)
Table of Contents
