Cell-free Massive MIMO and Radio Stripes

I have recorded a popular science video that explains how a cell-free network architecture can provide major performance improvements over 5G cellular networks, and why radio stripes is a promising way to implement it:

If you want more technical details, I recommend our recent survey paper “Ubiquitous Cell-Free Massive MIMO Communications“. One of the authors, Dr. Hien Quoc Ngo at Queen’s University Belfast, has created a blog about Cell-free Massive MIMO. In particular, it contains a list of papers on the topic and links to the programming code of some of them.

30 thoughts on “Cell-free Massive MIMO and Radio Stripes”

  1. Hi Emil,

    Thanks for sharing.
    Any comments on where the PAs sit for these antennas? What type of processing can be done in these black boxes?


    1. The PAs are placed next to each antenna. I’m imagining an output power per antenna of around 0.1 W, similar to a mobile phone.

      As much processing as possible should be done in the black boxes, so that the complexity of the tasks that must be done centrally becomes independent of the number of antennas – this gives a kind of scalability for systems with massive number of distributed antennas. In the uplink, it could, for example, be MR or LMMSE processing using the locally available CSI. The softly decoded data signals are then accumulated along the stripe so that the central processing unit only gets an accumulated version that is used for final hard decoding.

  2. Do you think the power consumption is a big problem, as all active components are placed on the plastic stripes?

    1. This will of course be taken into account in the design. Suppose the stripe is using a design similar to an Ethernet cable with a power over Ethernet protocol. Then we could use at least 100 W per stripe, which would be enough for 100 antennas (0.1 W transmit power per antenna and the rest for antenna processing).

      1. Prof. Emil

        Thank you so much for the reply.
        I noticed that the antennna should be located along some straight line, as shown in your paper, so is it because the blended antenna’s performance will be degraded and disabled, including gain and directivity diagram?

        1. The paper only contains illustrations and are not conclusive when it comes to deployment strategies. One of the good properties with radio stripes is that they can be bended and deployed in many different ways. If the antennas are roughly omni-directional, there is no need to think about directivity when deploying the stripe. If the antenna elements have clear directivity, one probably needs to have antennas that point in different directions or small arrays with beamforming capability to ensure coverage in all directions.

  3. Hi Emil,

    I’m interested in the 5g vs cell free massive MIMO SINR diagram at 20:00 mins. In 5g SINR color diagram you have given different cones with less yellow color (good SINR part). Each cone represent a beam (I mean ssb indicated beam) in 5g? (Question 1). If so good SINR part is less, so you mean even though we do hybrid beamforming for a UE, within the beam there will be cell center, middle and edge kind of SINR variations? (Question 2). My basic question is in 5g the hybrid beamforming done for a cell center vs cell middle vs cell edge UE varies? I mean SINR seen by the UE will varies? Please let me know.

    1. Question 1: No, each cone in the cellular case represents a cell and the data rate that a user can get at different places in the cell.

      Question 2: Since the base stations are well separated in a cellular network, there are large variations in signal strength between different parts of the cell (center, middle, edge). It can very by 1 million times! Beamforming improves the signal quality for all UEs but it does not reduce the signal strength variations.

      PS. I’m not sure why you are asking about _hybrid_ beamforming. It is digital beamforming that the system should preferably use.

  4. Hi,
    Thanks for sharing. I have one question, what is the stochastic channel model for cell free massive MIMO? In other words what are the differences between channel model in cellular massive MIMO and cell free massive MIMO?
    Thanks in advance

    1. There are no major differences. Correlated Rayleigh fading models are often used in both cases. 3GPP has propagation models (pathloss + shadow fading) that are designed for indoor and microcell usage. Those models are appropriate also for cell-free massive MIMO. A cell-free system might be deployed in the same way as a cellular network with small cells. The major difference is that users are served by many access points, instead of only one.

  5. Hello Professor Bjornson.
    I have a question.
    In cell-free systems which channel is more preferable channel with correlated Rayleigh fading or channel with uncorrelated Rayleigh fading? Is not it depended on the number of APs antennas?

    1. I don’t think there is a clear answer to this yet, but my impression is that uncorrelated fading is preferable since every AP has so few antennas that it cannot make efficient use of the spatial correlation.

  6. Hello professor Bjornson
    Thank you for your useful answers, as always.
    As I know in cellular massive MIMO systems the pilot signals are transmitted in UL for channel estimation and that is a principal for the system.
    I want to know is there such a principal in the cell-free systems?
    Can not we achieve a better performance if we transmit the pilot signal in downlink in cell-free systems?

  7. Hello Professor Bjornson.
    I have two question about one of your paper “Multiple antenna for beyond 5G”
    1) You say in somewhere of the paper that in cell-free systems all APs only need to know the CSI between themselves and all the users and do not require to share the CSI. I want to know whether this statement is always true for any of receive combining vectors?
    2) According to the figure 3 in the paper, to some extent you think power control is essential in cell-free systems? Is not it benificial for all users to transmitt with full power?

    1. 1) Well, you can certainly create combining schemes where the statement is true. But there are normally two categories: Distributed schemes where each AP processes the received signals using its local CSI, Centralized schemes where each AP sends it local CSI to the CPU which performs the processing. In both cases, each AP only has local CSI.

      2) Power control is always important, but my impression is that it is harder to perform downlink power allocation than uplink power control. Transmitting with full power in the uplink seems to be close to the sum-rate maximization solution.

  8. Hi, Professor Bjornson.
    As you have mentioned in section 7 in your book, we have some optimization criteria for power control that you have pointed out them over there, such as max min SINR.
    A question that arises to my mind is that these criteria are used for all systems?
    Can not we have other criteria for power control?
    For example in Cell-Free systems, because of the some different aspects of the system can we have other criteria for power control different from criteria that you have mentioned in your book.

    1. You can choose any criteria that you like. There is no optimal criterion but it is up to the system designer to pick something that he/she believes will lead to a suitable tradeoff between sum throughput and fairness.

      Note that even if the same criteria can be used in many different types of systems, the solutions that maximize the criteria will be different. I have an upcoming book where describe more about this: https://www.nowpublishers.com/article/Details/SIG-109

  9. Hi, Professor Bjornson.
    If I truly understood, in a user-centric network the users should be able to select the best servicing APs in each moment.
    I think this approach maybe need more control signals than a network with network-centric approach and therefore more overhead. This may be considered as a limiting factor.
    What is your opinion? Do you agree?

  10. Hi, Professor Bjornson.
    Thanks for your answers and the references you recommended.
    If we imagine a scenario in which we have a mobile user in a user-centric and also in a network-centric approach.
    I guess that because in user-centric approach the mobile user probably should select its serving APs faster than in a network-centric approach, so this might do with more control signals.
    Can I ask what are the main factors affecting the amount of control signaling?

    1. In both cases, the user device will connect to one AP in the access phase. In the network-centric approach, it is predetermined that if you connect to AP X, you will always be served by the predefined cluster of APs that AP X belongs to – even if that set is not involving all of the APs that surrounds you. In the user-centric approach, the cluster of APs that will serve the user is determined to involve all the surrounding APs. Some additional signaling might be needed in the access phase to form the cluster, but I don’t think it will make a major impact. It is rather an architectural question: Is the network built in such a way that cluster can be formed dynamically based on the user needs. In a network having one or multiple edge-clouds that carry out the processing (called central processing units in Cell-free Massive MIMO), this can easily be done. See for example this paper: https://arxiv.org/abs/1902.11275

  11. Hi, Professor Bjornson.
    As we know the channel hardening in cellular massive MIMO is an advantage which help us to design a power control algorithm only based on large scale fading. But in cell-free systems the channel hardening is weaker than the cellular massive MIMO.
    I want to know
    1) whether in cell-free systems, is it sufficient to design a power control algorithm only based on large scale fading.
    2) if we write an algorithm based on small scale fading, the algorithm should be run in each coherence block. Can we design such an algorithm. I think the algorithm may need more time for run than a coherence block especially in scenarios with high mobility.

    1. 1) Yes, I think it is usually sufficient (since there are many more access points than users in cell-free), but one can certainly create scenarios where this isn’t the case.

      2) This comes down to how efficient the algorithm (e.g., in terms of flops) and how efficient the implementation is (e.g., dedicated circuit vs. general purpose processor). Of course, we will be limited by the coherence time in practice. Deep learning can sometimes be utilized to shorten run times: https://arxiv.org/pdf/1901.03620.pdf

    1. Each user can be at a random geographical location in the network, and will be subject to interference from randomly located co-users. This leads to random variations in what kind of data rate (or spectral efficiency) that a user will get. We often describe the user performance in terms of the CDF of that randomness, so that one can see the performance variations.

  12. Dear Prof. Björnson,

    I am a master’s student at Politecnico di Milano studying in this field. I have some questions about this technology.
    How is the fronthaul capacity calculated, and which parameters depend on the fronthaul capacity such as coherence time, Modulation and Coding Scheme, quantization of the ADCs, active user equipment quantity, antenna quantity, etc?
    What is the bit rate typically?
    I will be appreciative if you reply to my questions. Thanks in advance.

    1. The answer depends on the functional split that is used; what is done at the AP versus the cloud. One option is:

      Symbol rate (bandwidth) * number of spatial streams * bit per symbol

      The bit per symbol varies with the modulation and coding, but you need to dimension the fronthaul for maximum load.

Leave a Reply

Your email address will not be published. Required fields are marked *