Beamforming for MU-MIMO in 5G New Radio
Overview
In this webinar, you will learn about single- and multi-user MIMO in 5G NR, as well as common beamforming techniques and scenarios. We will cover different techniques to estimate the channel or channel information for beamforming purpose, how the UE communicates the necessary information back to the gNodeB, the different options for SU-MIMO vs MU-MIMO, and how to simulate a complete link with adaptive beamforming in MATLAB with 5G Toolbox.
Highlights
- When to use SU-MIMO vs MU-MIMO
- Relationship between channel estimation and beamforming
- The use of channel state information reference signals (CSI-RS) for beamforming purposes or beam refinement
- Types of codebook feedback (Type I, Type II, eType II) and application to SU/MU-MIMO
- Sounding reference signals (SRS) based beamforming on reciprocal channels (for TDD scenarios)
About the Presenter
Marc Barberis is an application engineer with MathWorks, based in Silicon Valley, where he focuses on wireless standards such as 5G NR, LTE and WLAN. Prior to joining MathWorks, Marc worked on the modeling of telecommunications systems and the design of receivers for 2G/3G cellular phones and satellites systems among others. He attended numerous 3GPP standardization meetings for GSM/EDGE and W-CDMA. He holds a M.S. in Signal Processing from Ecole Nationale Superieure des Telecommunications de Paris, France.
Recorded: 25 Jul 2023
Hello and welcome to this presentation of Beamforming for multi-user MIMO in 5G New Radio. My name is Marc Barberis. And I'm part of the application engineering group at MathWorks, where I focus on our wireless standard products, including 5G, LTE, and wireless Lan. I have been with MathWorks for many years.
And at MathWorks, and with prior companies, I have followed the different cellular standards from GSM through 5G, including all flavors of 2G, and 3G, and then later, 4G. Today, we are going to talk about beamforming in 5G at a high level before diving into some of the tools that the standard provides to enable beamforming.
First up, sounding reference signals then channel state information reference signals, or CSI-RS. At that point, we have a dedicated section about codebook design, on which CSI feedback relies. In the last section, we'll discuss advantages and disadvantages of these different techniques for single and multi-user MIMO.
Let's start with a quick overview of beamforming in 5G and basic concepts associated with beamforming. Beamforming is certainly a key to 5G NR. Both data and control channels are being formed with their own sets of dedicated reference signals using the same beamforming to make it transparent to the receiver.
Beamforming is usually characterized as digital, analog, or hybrid. While the details vary depending on the implementation and the frequency range, both digital and analog beamforming are present in 5G devices. In this presentation, we'll turn our attention towards digital beamforming. I added a few links at the end of this presentation to point you to resources for analog and hybrid beamforming.
Another way to classify beamforming is between open loop and closed loop schemes. For example, beam sweeping which consists in following a set pattern to transmit a beam in different directions to cover a sector is an open loop scheme. Other schemes, such as codebook based beamforming are closed loop because they require feedback from UE the base station to determine a suitable downlink beam.
In this presentation, we'll be focusing on codebook and non-codebook based beamforming. You can look at the links on the right for examples of beam sweeping and beam refinement. If you want to target the receiver most efficiently, you need to know how to best beamform the waveform. And for that, you need information about the location of a receiver, or more generally speaking, about the propagation channel between the transmitter and the receiver.
To obtain that information, there are two options. Either you can measure it based on a known signal you're receiving. Or you have to get that information fed back from the other side of the transmission. That other side can estimate that information based on a known signal you're transmitting. How does that look like in 5G?
First, let's look at downlink beamforming. The base station can estimate the uplink channel based on the sounding reference signals which the standard provides. Using the uplink measurement to target the downlink only works when the channel is the same in both directions, which means that both transmissions occur on the same frequency at different times. That is the principle of Time Division Duplex, or TDD.
Or the base station can ask the UE to report information about the channel based on measurements performed on the channel state information reference signals. This approach works for both TDD and Frequency Division Duplex, that is FDD. We will look at both those options in detail in the following sections.
Similarly, on the uplink, the UE can estimate the downlink channel based on the CSI-RS and use that estimate to propose possible precoding for the uplink. This approach works only for TDD. Or the UE can receive precoding information directly from the base station. The base station can use the CSI-RS to estimate the channel and determine suitable uplink precoding.
At this point, and before diving into the detail of SRS and CSI-RS, let's quickly review why channel estimation is at the core of beamforming and how SVD, or Singular Value Decomposition, provides a reference for beamforming. The characteristic of OFDM systems such as 5G LTE, and wireless Lan is that transmission can be seen as affected by a flat fading for every subcarrier.
This simplifies equalization and channel estimation because the channel between one transmit and one receive antenna is a complex number with phase and amplitude information for each subcarrier. This means that when considering a single subcarrier, the channel can be summarized as a matrix H of dimension the number of transmit by the number of receive antennas. And the received waveform is simply H multiply by X, where X is the vector of transmitted values from each antenna.
For any matrix H, we are guaranteed a singular value decomposition of H as the product of U, S, and V prime. Where U and V are unitary matrices and S is a diagonal matrix. The diagonal elements of S are called singular values. Why am I bringing the SVD? That's because we can now rewrite the received vector as Y equal U, S, V prime X. And if we decide to transmit V multiplied by X instead of X, the received vector then becomes U multiplied by X multiplied by X because V is a unitary matrix.
So what does transmitting V multiplied by X mean? Let's write out the V matrix and look at the first column. That column is a precoding vector because what it does is take the first value of X, which is the first layer, and map it to N antennas with a given amplitude and phase relationship. And the same applies to all other columns of V, which represent additional precoding vectors for the other layers.
As we just saw, using V as a beamformer results in each layer being multiplied by a singular value at the receiver ignoring the multiplication with unitary matrix U. Actually in practice, the base station chooses to transmit only as many layers as there are singular values with decent energy. So it only uses the first L columns of the V matrix. So the V matrix from SVD decomposition of a channel is a great estimate for a beamforming matrix.
Now, let's have a look at the sounding reference signals. Sounding reference signals are provided for different use cases. Those include beam management, for example, to assist the base station in determining a suitable receive beam. For that purpose, the same SRS can be sent multiple times with identical beamforming to let the g node Be try out different receive beams and pick the best one.
SRS Can also be used for to UE to propose a precoder for the uplink prior to applying it to the data channel. The base station can then allow or disallow particular precoding vectors in the proposed beamforming matrix. Here, we will focus on the other two cases, where the SRS is used as a means to estimate the uplink channel.
This can be to propose uplink beamforming, known as codebook based transmission, or to determine suitable downlink beamforming in the case of a reciprocal channel. This case has the strange name of antenna switching in the standard. And we will explain later where this name is coming from. Finally, note that there are other use cases such as positioning, which was introduced in release 16 of the standard.
Let's first look at the structure of the SRS. This slide summarizes some of the main points in terms of antenna ports, length, OFDM symbols, comb structure, and bandwidth. We'll look at those in more detail in the next slide as a picture is worth a words. This is the time frequency structure for the SRS.
It is transmitted over one or several symbols, marked in light blue. Within these symbols, it occupies a comb structure every two, four, or eight resource elements. Here, we picked four, as shown in dark blue. The location of those SRS symbols is anywhere within the slot with consecutive symbols assigned in case there are more than one.
Finally, the bandwidth can span between 4 and 272 resource blocks. There is a table in the standard that lists all possible configurations, including the bandwidth. Over time, the UE can transmit the SRS periodically every n slots. There is also a mechanism to pause and restart transmission. This is called semi-persistent scheduling.
Finally, it is also possible for the base station to request a one-time SRS transmission. Such a transmission will be triggered by a DCI, or Downlink Control Information, transmitted on the downlink control channel. When considering SRS, it is important to realize that multiple UEs can transmit SRS in the same slot. SRS transmissions from different UEs can be separated by time, which means a different OFDM symbol, frequency, meaning a different comb, and/or a cyclic shift of the underlying sequence.
In this case, the two SRS overlap in time and frequency but are orthogonal because of the cyclic shift. Let's have a look at some of the challenges associated with SRS. The first and critical one is that you UEs are power limited. This means that it is not possible for a UE that is far from the base station to transmit over a large bandwidth and be received with sufficient SNR by the base station to perform channel estimation.
Another factor to consider is that many UEs may have fewer transmit antennas than receive antennas. The issue here is that the UE may not have one separate RF chain for each antenna on the device. It can receive from, say, four antennas, but only transmit from one at a time. Such a UE is called 1T4R.
Using SRS to estimate downlink beamforming relies on the fact that in TDD, the propagation channel is reciprocal. As we will see later, things are never as simple as they appear. And channels may not be as symmetrical as we would like them to be. Finally, carrier aggregation can increase the challenge for uplink sounding as the base station may use downlink aggregation, which means the UE would have to transmit SRSes in all carrier components.
Let's tackle the first issue, limited transmit power. The UE can trade bandwidth for power. It can transmit at a higher power over a smaller bandwidth. Therefore, the standard offers frequency hopping as a means to cover the whole bandwidth over several transmissions. In addition, the UE can repeat the SRS in order to increase the SNIR, the base station, and extend the transmission range.
Here is an example of frequency hopping. Here, the bandwidth in is divided into two. And the UE transmits with full power over half the bandwidth in two different OFDM symbols. This example combines frequency hopping and repetition. Here, the received power increases by 4x because each transmission can be twice as high and is repeated twice. This hopping and repetition can take place in different slots, leading to interslot frequency hopping.
While frequency hopping and repetition help compensate for limited transmit power, the downside of this strategy is that it takes longer for the base station to acquire a view of the complete channel. This is how you can set up actual examples of frequency hopping and repetition with MathWorks 5G Waveform Generator, part of 5G Toolbox.
Here, we have the 5G Waveform Generator. And we have set up SRS transmission and disabled data transmission. The CRS and BRS values correspond to an SRS bandwidth of 48 resource blocks, as defined in the table in the 5G standard. We are using two OFDM symbols with no repetition.
We can see intraslot frequency hopping, along with interslot frequency hopping. If we set the repetition to two, both OFDM symbols within a slot are used to cover the same resource blocks. And we only have interslot frequency hopping. If we increase the number of OFDM symbols to four, then we get a similar pattern to the original one. But each transmission is repeated on two consecutive symbols within one slot.
If I show the detail of the SRS allocation within one resource block, we can see the comb structure with 1 in 4 resource elements carrying SRS. If I set that parameter to two, the SRS density doubles. Let's generate the waveform. We can see that the SRS covers the bottom 144 resource blocks of the bandwidth part with this setup. And we can see a summary of those parameters on this slide.
After discussing how to remedy limited UE transmit power with frequency hopping and repetition, let us now talk about the case where the UE has more receive and transmit antennas. There is a mode in the specifications, antenna switching, which directly refers to this case. The issue at hand is that the base station may need to know the channel to all receive antennas in order to use the right precoding on the downlink, especially to send multiple layers.
But the UE cannot transmit for more than one antenna at a time. In this case, the standard foresees the possibility to transmit from each antenna in succession. Whereby the UE dynamically switches which antenna is connected to the RF transmit chain. Here is what it looks like.
Antenna one transmits an SRS. Then antenna two transmits with a gap provided for the UE to perform antenna switching. Note that the name antenna switching in the standard now denotes this case where the SRS is used to assess the channel for downlink beamforming, irrespective of whether the UE has fewer transmit antennas or not.
I mentioned earlier that even in TDD, the channel may not be reciprocal. We have just seen one example of asymmetry in that there may be fewer transmit than receive antennas. Still dealing with the RF chain, there is an even greater source of asymmetry. The receiver cannot separate the propagation channel which is reciprocal from the RF chains that the waveform passes through.
For the receiver, the channel is the concatenation of the transmitter chain, the propagation channel, and the receiver RF chain. There is no reason why those uplink and downlink chains with vastly different power and complexity constraints would behave identically. The base station can have much costlier components. But it also transmits at a much higher power.
Here is an example in 5G Toolbox which demonstrates the use of SRS for downlink beamforming. This is a very sophisticated 5G example that shows how to perform multi-user downlink beamforming based on channel estimation performed on the reception of SRS sent by multiple UEs. The base station determines not only the beamforming matrix, but also which UE to schedule based on the quality of the channel for each UE and their orthogonality. And the scheduling happens on a subband basis. Let's run the example and look at the scheduling assignments.
Scheduling takes place after each SRS transmission. This picture shows which resource blocks are assigned to which UEs. The scheduler is using multi-user MIMO. As we can see that the same resource blocks are assigned to multiple UEs. Can also see that depending on the channel conditions, the assignments change.
Now, let's have a look at the channel state information reference signals. Channel state information reference signals are multifaceted animals. And some of them don't even exist. Or rather, they have zero power. CSI-RS are used for beam management, interference measurement, and channel measurement, among other. Here, we will focus on channel measurement.
But first, a quick word about zero power CSI-RS for interference management. Those correspond to locations that the base station reserves to have no power. This enables the UE to measure the inter-cell interference level, as the only energy at those locations comes from surrounding base stations.
CSI-RS are always confined to a bandwidth part, but contrarily to DMRS. They are not limited to the bandwidth of current data transmission if there is data transmission. Like for SRS scheduling, there are periodic, semi-periodic, and aperiodic transmissions, whereby the period for CSI-RS can be between 4 and 640 slots.
The frequency density of CSI-RS is either 1 or 1/2, meaning that there can be CSI-RS in every resource block or in every other resource block. The number of resource elements allocated depends on the CSI-RS configuration. These are the four blue squares here. And this is what we want to look at on the next slide.
A CSI-RS configuration can specify up to 32 ports. This means that there can be 32 orthogonal transmissions on 32 ports. Like for SRS, CSI-RS can be separated in time, frequency, or by sequence orthogonality. The time and frequency separation correspond to different colors on this slide.
In addition, when several ports use the same time frequency resources, which means the same resource elements shown in the same color, they can be separated by frequency and/or time, Code Division Multiplex, or CDM. For example, if you look at the case with eight ports, you can see two different sets of resources in blue and orange.
Each one of them occupies four resource elements. And to each port corresponds one of the four cover codes shown with pluses and minus 1's. This amounts to two times four possible orthogonal courses, one for each one of the eight ports supported. The same applies to all configurations with 2, 4, 8, 12, 16, 24, or 32 ports.
This is how you can set up a configuration with MathWorks 5G Waveform Generator, part of 5G Toolbox. Here, we have the 5G Waveform Generator. And we have set up CSI-RS transmissions as well as the synchronization signal burst in light blue. And disabled data and control for clarity.
The row value of eight corresponds to an eight port configuration, as per the 5G standard. When we turn on the resource element mapping visualization, we can clearly see the four resource elements associated with the CSI-RS on this port. And if we generate the waveform, we can see the part of the spectrum corresponding to CSI-RS transmission on the left as well as the synchronization signal block in the middle. We can see a summary of this configuration on this slide.
It is time now to consider what the UE does upon receiving the CSI-RS. The main objective that we are focusing on here is channel measurement for reporting to the base station. But once the channel has been measured, it turns out that the channel estimate would be much too large to feed back to the base station as is. For that reason, and this is what is going to keep us busy in the next section, the standard proposed is one way to reduce the amount of feedback by selecting one precoder from a codebook of matrices.
Or rather, the standard proposes three different types of codebooks for different situations. And again, this is the focus of the next section. Before that, note that the gNB may opt to select a different precoder than the one proposed by the UE. And one possible reason can be to minimize interference to other UEs.
So before looking at the codebooks in detail, the final piece I wanted to mention here is the concept of CSI-RS resource index. This index can be reported by the UE. It is useful when the UE receives several CSI-RS resource sets with, for example, different beams applied. This lets the UE identify which resource set the CSI report is based on. On this picture, it would identify resource set one and then send a CSI report about that resource set.
This leads me to the next part of this presentation, where we discuss the codebooks that are defined in the standard. It may seem strange to start this discussion about codebooks with a view of the propagation channels. But channels are the heart of the decisions made for codebooks and reporting.
You can look at a propagation channel in time and frequency. But you can also look at it in space. 3GPP defines several channel models called Cluster Delay Lines, or CDL as representative of conditions that will be experienced in 5G deployments. Some of the profiles are non-line-of-sight, that includes profiles A, B, and C. The other ones, D and E, are line of sight.
What we're trying to see here is what kind of spatial energy distribution we can expect with those channel models. Switching to MATLAB, here, I have a simple piece of code where I define a CDL-C channel model. And receive array, that lets me pick the energy from any direction I want. I transmit constant ones through the channel. And I iterate on the direction of the receiver array to represent the energy received at a function of the azimuth and elevation.
As we can see, most of the energy is concentrated in the spatial domain. If I rotate the view to look at it from the top, it is quite striking how sparse the channel is spatially. This is even more striking in the case of a line-of-sight channel, such as CDL-D, as we can see now. This tells us there must be something we can do to compress that spatial information.
So we have seen that the channel information reported by the UE to the gNodeB must be compressed. And we are going to do so in three ways. Time, by periodically reporting new information. Frequency, by providing either wideband or more detailed subband information. And space, only a few directions carry information. And this is what the standard exploits in codebook design.
Codebooks were designed to efficiently report spatial and frequency information. Three types are present in release 17 of the standard. Type 1, type 2, and enhanced type 2, or E type 2. In addition to the assumptions about the type of propagation channels, one important assumption made in designing the codebook is the type of antenna array. They are assumed to be linear or rectangular arrays, possibly multi panels with dual cross-polarized antennas.
If we look at the standard, we'll see tables like the one on the right, which specify different possible configurations and associate a number of CSI-RS ports with numbers N1 and N2. Those numbers correspond to the form factor of the antenna. For example, the first line of the 16 antenna ports shows 4 and 2 for N1 and N2, meaning that the underlying assumption for the base station antenna is that it has a 4 by 2 structure.
And each one of the eight antenna locations is comprised of two antennas polarized in a different direction. This means that the total number of antennas is indeed 2 multiplied by N1 by N2, or 16. This cross-polarization will be quite apparent in the structure of the codebooks, as we will see on the following slides.
The basic idea of the codebook is to exploit the DFT, which is computationally efficient. The precoding vectors in the codebook are DFT-based. And each one of them corresponds to a different beamforming direction. As it turns out, using a set of DFT beams would lead to two coarse a quantization of the beam direction, which is why the standard uses a 4x oversampling in both directions to produce a finer representation of directivity.
On this slide, I'm showing the basic formula for beamforming with a Uniform Linear Array, or ULA. A sine wave impinging on a ULA at angle theta arrives at the second antenna with a slight delay tau, which is proportional to sine theta and the distance between the antennas. If we represent the phase shift of the sine wave between antennas one, two, and three, you get the formulas on the right. I want you to keep those in mind, as we are going to compare them with what happens with DFT on the next slide.
If we consider a one DDFT, the coefficients of the summations are represented under each antenna. And they constitute a DFT vector. If we compare those coefficients with the formula from the previous slide, we can see that the DFT vector is indeed a beamforming vector in a particular direction.
So now that we have those arrows in our quiver, let us look at the fundamental formula for codebook type 1, W equal W1 multiplied by W2. This formula is not explicitly stated in the standard. But it represents how the codebook is built. W1 is a block diagonal matrix with matrix B on the diagonal. B is comprised of one DFT vector of length N1 by N2.
Each matrix B applies to a different polarization of the antennas. This means that both polarizations use the same beam, which is an important fact to note. W2 represent phases between polarized antennas. Taking the simplest example of a one layer on the right, W2 is a two element vector, 1 and phi.
This results in W being the DFT vectors repeated with a different phase for the second polarization to make it a 2 multiplied by N1 multiplied by N2 vector. The UE reports the numbers L and M, as well as the angle phi through index N. This is a view which I hope is helpful of the tip or main lobe of each one of the possible DFT vectors that underlies the codebook.
The UE reports the index of the DFT vector that it picked as the best precoder among the orthogonal DFT vectors and any of the beams from the 4x oversampling. This beam is the dark blue one. The next question is how it should report the best beam for other layers in the case of multiple layer transmission.
Letting the UE pick any beam from any layer would be ideal from a performance standpoint. But it leads to too many bits to report. Therefore, the standard restricts the location of beams for subsequent layers. The second layer can only use the same beam or one of the other three beams shown in orange. And this way, only two bits are necessary for signaling. The phase is even more restricted with only two possibilities, meaning one bit.
For more layers, even stronger restrictions come into play. In the case of six layers, the relative positions of the beams for all six layers are predefined. An important aspect of codebooks is the recognition that matrices W1 and W2 are different in nature. W1 includes the DFT vectors. They represent a general direction. That direction is expected to be the same across the bandwidth and to only vary slowly over time.
W2 represents the cophasing of the cross-polarized antennas. It accounts for fast fading. And it varies over the bandwidth. Furthermore, it varies faster over time. As a result, W2 is more costly because it often has to be reported per subband and is in need of more frequent updates.
This also explains why the standard defines several types of reports, including full and partial reports. A coarse report typically with just W1 is smaller in size and can be transmitted over control or PUCCH. While a complete report, including subband information requires more bits and will be sent over the data channel, or PUSCH.
Summarizing the pros and cons of codebook type 1, on the pro side, codebook type 1 supports up to eight layers, the maximum number defined in the standard. Even though I just said W2 is large, it isn't that large because of the restrictions mentioned earlier. On the other hand, each layer can only be associated with a single beam. The definition for phase relationship is QPSK, two bits. And there are all those restrictions just mentioned. As a result, it is generally considered that codebook type 1 is too coarse for use with multi-user MIMO, where transmissions to multiple users require better targeting.
This leads naturally to wonder what is different with codebook type 2. First, each layer can select up to four beams and not just one. As there are multiple beams, the report now includes the relative amplitude of those beams with eight levels, meaning three bits. In addition, the relative beam amplitude between subbands can be reported with two bits per subband. And the phase for strong beams can use three bits. This scheme only supports one or two layers per UE. But it provides a more accurate view of the channel at the cost of increased feedback.
Here is a closer view of codebook type 2. We're not trying to understand all the detail. But I want to make a few observations as they highlight what I've just mentioned. First, the precoding matrix is composed of two vectors. Each one represented by one summation.
As for codebook type 1, they correspond to each polarization of the antennas, the first one on the first line and the second one on the second line. The summation is over the number of beams per layer selected, something not present in codebook type 1, where only one beam is selected per layer. There are two parts to the summation, the orange wideband part and the blue subband part.
The first term in orange is the DFT vector. They do not depend on the layer l, which means that both layers, if there are two of them, use the same vectors. The second part is the set of relative amplitude for each beam and each layer. The second term in blue is comprised of the relative amplitude between subbands and the phases per subband.
So it is natural at this point to wonder how codebook type 1 and type 2 compare. While there is no general answer, the example here, built with 5G Toolbox, illustrates the difference for a simple case of single layer transmission. We define a simple propagation channel with two paths between the gNodeB and the UE. The upper one having a stronger amplitude, as represented by the size of the circle at the reflector.
Let's have a look at the example in Matlab. I am varying the angle of a second smaller reflector between minus 20 and minus 5 degrees. And for each configuration, computing the beamforming vector based on codebooks type 1 and codebooks type 2. We can see that codebook type 1, in blue, always picks the same beam, which targets the strongest path.
The pattern shows side lobes that just belong to that DFT vector. In contrast, we can see that for angles minus 20 and minus 15, codebook type 2 clearly takes advantage of the secondary reflection by directing a secondary beam in that direction. In the case of 5 degrees, it picks the same solution as codebook type 1, which further shows that results are not always intuitive or easy to interpret.
The shortcomings of codebook type 2 are the limitation to two layers, the fact that both layers must use the same beams, and importantly, the high payload required to transmit this more detailed report, especially when including subband amplitude information. Keep in mind that the subband matrix W includes phase and amplitude information per beam and per subband.
So in the spirit of continuous improvement, 3GPP introduced codebook eType 2 in release 16 of the standard in order to alleviate some of the shortcomings of codebook type 2. First on the list, support of up to four layers without increasing the payload. The focus of eType 2 is to target the frequency domain reporting. In type 2, most of the bits are required because of subband reporting.
Even though there are relative quantities between subbands, most efficient compression techniques can easily be devised. If you're familiar with image compression, you know that reporting the difference between two lines in an image is nowhere near optimum. JPEG uses DFT to compress that information. And that is what happens with eType 2, where DFT compression is used to represent variations in the frequency domain.
Another way to look at it is that in a sense, selecting a few DFT points in frequency is a similar concept to selecting a few beams in space. I will not get into more detail about eType 2 in this presentation.
Before concluding this discussion about codebook design, I want to point out two examples from 5G Toolbox which illustrate what we have just discussed. The first one computes channel quality information, precoding matrix information, and rank information for type 1, type 2, or eType 2 codebooks. This example is what I use as a starting point in the example we just saw illustrating the difference in beams between codebooks type 1 and 2 for a single layer.
The second example builds on this reporting to dynamically adjust the transmission parameters for a 5G link, including the modulation encoding scheme, beamforming matrix, and number of layers transmitted. It puts everything together and measures the data throughput with a CSI closed loop scheme. I will get back to this example in a moment. These are very advanced examples that can help you assess the impact of various codebooks.
As hard as 3GPP worked to design those codebooks, it turns out that performance with any of these code books still significantly lags performance that could be achieved if the gNodeB had perfect knowledge of the channel. This is why this whole area of codebook design is being looked at as we speak in release 18 developments as a good candidate for AI-based feedback. The idea here is to use AI to find a more efficient way to compress the channel information.
One proposal consists in taking the channel estimate, which is a large multi-dimensional variable with dimensions for subcarriers, OFDM symbols, slots, all transmit, and all receive antennas, and preprocess it to start reducing its size. For example, averaging over time, and DFT-based compression. At that point, AI can take over to efficiently encode the reduced information and decode it on the other side with as small a loss as possible.
Such a process is precisely what you will find in the CSI feedback with autoencoder example in 5G Toolbox. The example generates training and test data, trains the AI network, and analyzes its performance, including under various quantization constraints. This is a great example to look beyond code books and understand how to apply AI to wireless applications.
So let me get back briefly to the example we saw five slides ago. I want to point out that this example also incorporates the ability to introduce AI-based compression into the throughput simulation. That is in addition to the traditional RI-PMI-CQI method. And this lets you compare performance of traditional or AI-based methods to perfect CSI feedback.
To wrap up this presentation, we are going to compare the use of CSI-RS and SRS, as well as single-user MIMO versus multi-user MIMO. Regarding SRS versus CSI-RS, let's start by stating the obvious, which is that both methods for channel estimation work very well when the UE is stationary and close to the gNodeB.
When the channel is reciprocal, which means in TDD, you do have a choice as to whether to use SRS or CSI-RS for uplink and downlink beamforming. The main advantage of using reciprocity is that there is no need to feed back and compress information. As we've just gone through codebook design and associated challenges, this seems very attractive.
So what is there not to like about SRS for downlink beamforming? Well, as we saw earlier, UE power is limited. Repetition and frequency hopping can help, but at the expense of feedback delay, which equates to what is called channel aging. By the time the information about the channel has been collected, it is out of date because the UE moved.
Therefore, in some deployments, it may be found that using CSI-RS for UEs that are further from the base station can be advantageous, especially if the base station beamforms the CSI-RS to increase their range. Finally, let's talk a little bit about the use of multi-user MIMO.
Setting up multi-user MIMO requires detailed channel information, which requires, as we have seen in our codebook design section for downlink beamforming, a large payload. In addition, multi-user MIMO is best suited to cases with multiple UEs with high SINR Those UEs should also have reasonably high throughput requirements. Otherwise, the overhead of setting up multi-user MIMO transmission is not worth it.
In practice, having multiple UEs in a cell with high traffic is rarely found in an environment where neighboring cells have low traffic and low number of UEs. This means that it is often the case that those favorable conditions for multi-user MIMO do not come along with a high SINR situation.
A suitable SINR would be more frequently obtained in a cell with many medium low traffic UEs. And in that configuration, setting up multi-user MIMO is not necessarily worth it because of the overhead or simply the delay required in setting it up. For those reasons, it turns out that for all the emphasis on multi-user MIMO, single-user MIMO can give better performance on the downlink.
On the other hand, a maybe surprising place where multi-user MIMO can make sense is on the uplink. Often, UEs cannot transmit efficiently more than one layer because of power limitations. In such a case, coscheduling several UEs with single layer each on the uplink may prove an efficient use of frequency resources.
This brings me to my final slide, where I wanted to provide a few links to examples and related webinars, as well as the 5G Toolbox page. The first three examples in the list are the ones I mentioned in my presentation. This concludes my presentation. I hope you found something of interest. The examples and capabilities mentioned in this presentation are current as of release 2023a.