Beyond the Cellular Paradigm: Cell-Free Architectures with Radio Stripes

I just finished giving an IEEE Future Networks Webinar on the topic of Cell-free Massive MIMO and radio stripes. The webinar is more technical than my previous popular-science video on the topic, but it can anyway be considered an overview on the basics and the implementation of the technology using radio stripes.

If you missed the chance to view the webinar live, you can access the recording and slides afterwards by following this link. The recording contains 42 minutes of presentation and 18 minutes of Q/A session. If your question was not answered during the session, please feel to ask it here on the blog instead.

Update: The recording from the webinar has been delayed (due to the virus crisis), so I have recorded an alternative video:

14 thoughts on “Beyond the Cellular Paradigm: Cell-Free Architectures with Radio Stripes”

  1. Hi Prof Emil,

    Thanks for the great presentation!

    My question is how do you see the potentials of machine learning applications in cell-free networks? I have seen your blog post and video on deep learning, but the machine learning that I want to stress here is the reinforcement learning rather than supervised/unsupervised ones.


    1. Thank you for attending the presentation!

      I think machine learning might be the key to implement cell-free networks in a distributed fashion. If you remember, there was a big performance gap on Slide 18 between the “heuristic” distributed implementation and the optimal centralized method. It doesn’t have to be like that, but it shows how hard it is to create man-made distributed algorithms. The sequential processing that I mentioned ( is one step towards reducing the gap, but I think that machine learning can provide the next step. Essentially, we need the APs to learn how to cooperate without having to send excessive amount of information between each other. Since we don’t know how to achieve that, reinforcement learning might be the right approach.

  2. Professor Björnson,

    Thank you for your presentation. I am still waiting for webinar recording to show up, but I had chance to look at your slides and your awesome popular science video.

    I am interested in experimental research (like channel measurements/radiation pattern change according to the material of the building structure) using radio stripes – do you by any chance know if there is any vendor that sells radio stripes/film antennas or are they still at the research/product development stage? Thank you in advance.

  3. Hello Professor,

    a cell-free massive MIMO (CF-mMIMO) system with K users and L access points (AP) is essentially, as far as I understand, a network MIMO system with L >> K that operates in TDD mode, employs coherent joint transmission (CJT) (hence a distributed mMIMO setup) , and adopts a user-centric architecture (implemented, for example, via the dynamic cell clustering (DCC) scheme), therefore appearing as “cell-less” or “cell-free” during transmission.

    Some related questions:
    1) As described above, CF-mMIMO is basically an evolution of the CJT introduced in coordinated multi-point (CoMP) that utilizes a massive number of APs and a user-centric architecture. However, in 3GPP Rel-15 and Rel-16 are also defined several multiple transmission / reception point (multi-TRP) variants, namely, CJT and non-coherent JT (NC-JT). The latter are divided into the fully-overlapped NC-JT and the non-fully-overlapped one. The former implies a centralized scheduler, fully-overlapped PRBs, and user data sharing. On the other hand, neither of these hold true for the latter (not even user data sharing (!), thus I really don’t understand why this is labeled as joint transmission, since each AP serves each own UE with each own data). My questions is: a) Are these NC-JT variants simply refer to CoMP-JT (coherent vs. non-coherent)? b) Under L >> K and a user-centric architecture, these NC-JT variants can be labelled still as CF-mMIMO or CJT is a prerequisite; c) If (b) is correct, are these approaches considered as a means to have a distributed implementation of CF-mMIMO (e.g., with local schedulers) and better scalability (e.g., with no data sharing)? Is this the reason why the main focus on Rel-15 and Rel-16 is on NC-JT instead of CJT?
    2) In general, a distributed implementation (or, in general, a scalable one) requires some enhanced processing capabilities at the APs?
    3) Does both cloud RAN and distributed RAN suit the needs of CF-mMIMO? For a centralized implementation, cloud RAN seem to be the optimal solution. However, I think that it may give rise to some scalability issues.

    I realize that this is a long answer, but I am confused regarding the jargon and details on the 3GPP Releases vs. the CF-mMIMO concept and its generalization / extension / evolution.

    1. I don’t know the 3GPP jargon either.

      1a) Yes, I believe it is the same thing.
      1b) CF-mMIMO is supposed to be based on coherent transmission. In a dense deployment, non-coherent transmission performs poorly in comparison. See for example Figure 7 in where we considered this.
      1c) I wouldn’t recommend non-coherent joint transmission. If one cannot afford coherent joint transmission, then I think that classical cellular operation will be roughly as good as non-coherent transmission.

      2) I think the main requirement is the phase-coherency between APs. Each AP will also be serving more users than in a cellular system, so this increases the complexity.

      3) I agree. A cloud-RAN implementation is preferable from a performance perspective, but it remains to develop a way to implement it in a scalable way. I think the sequential processing that we proposed in this paper might be a good tradeoff between centralized and distributed operation:

  4. I’m curious about UE cell acquisition. In cell-based mMIMO, during initial synchronization the channel can be enhanced without real-time CSI using fixed beam-scanning, because all transceivers use approximately the same channel to each UE. The channel can then be hardened with CSI. In a cell-less implementation, beam-scanning without real-time CSI does not seem so feasible during synchronization, at least in dynamic environment that features moving obstacles. Can you comment?

  5. What are the most challenging signal processing aspects for DMIMO or cell-free massive MIMO? What signal processing technique can I propose for a post-doctoral research? Please suggest me a very unique/novel research direction.

    1. You can find lists of open problems in the following papers:

      Jiayi Zhang, Member, Emil Björnson, Michail Matthaiou, Derrick Wing Kwan Ng, Hong Yang, David J. Love, “Prospective Multiple Antenna Technologies for Beyond 5G,” IEEE Journal on Selected Areas in Communications, To appear.

      Giovanni Interdonato, Emil Björnson, Hien Quoc Ngo, Pål Frenger, Erik G. Larsson, “Ubiquitous Cell-Free Massive MIMO Communications,” EURASIP Journal on Wireless Communications and Networking, vol. 2019, no. 197, 2019.

  6. Hi Prof. Emil,

    I read your paper titled “Cell-Free Massive MIMO With Radio Stripes and Sequential Uplink Processing” in ICC 2020 Workshops. I agree with you that radio stripe is an efficient implementation way for cell-free, and it is important to limit the front-haul loading in order to reduce the cost.

    However, I meet a small problem of reproducing the results. Two simulation results looks similar, but not the small. Therefore, I want to check the simulation settings with you to make sure I didn’t misunderstand your paper, including :(1) The user distribution. Is it a uniform distribution in a 125mx125m square? (2) Spectral efficiency is calculated using one user or the average value of multiple users. (3) ASD of the correlated channel.

    Thanks a lot.

    1. Hi,

      Thank you for the query.

      1. UEs are uniformly distributed:
      a. perSideLength = 125 m. Yes, setup was 125m x 125m.
      b. UEpositions = (0.1 + (0.9-0.1).*rand(K,1))*perSideLength + 1i*((0.1 + (0.9-0.1).*rand(K,1))* perSideLength);
      c. So leaving little area on all sides, UEs are uniformly distributed, this is done to avoid overlap of UE positions with APs.

      2. For SE: It’s SE per user. More clearly,
      a. For a given setup, SE is a vector with “k”th entry being SE for UE “k” which is averaged over realizations.
      b. All these vector SEs for all setups is concatenated and then CDF is plotted, so its SE per UE

      3. ASD used for simulation is 15 degrees.

      1. Many thanks. Really a quick and detailed response.

        The main differences turn out to be the UE distribution and definition of SE. With the alignment of these settings, I reproduce the simulation results with a TX power of 200mW, not 50 mW in this paper. I’ll check whether there is any calculation error in my program. Thanks again for your great job.

Leave a Reply

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