I have written several posts about Massive MIMO field trials during this year. A question that I often get in the comment field is: Have the industry built “real” reciprocity-based Massive MIMO systems, similar to what is described in my textbook, or is something different under the hood? My answer used to be “I don’t know” since the press releases are not providing such technical details.
The 5G standard supports many different modes of operation. When it comes to spatially multiplexing of users in the downlink, the way to configure the multi-user beamforming is of critical importance to control the inter-user interference. There are two main ways of doing that.
The first option is to let the users transmit pilot signals in the uplink and exploit the reciprocity between uplink and downlink to identify good downlink beams. This is the preferred operation from a theoretical perspective; if the base station has 64 transceivers, a single uplink pilot is enough to estimate the entire 64-dimensional channel. In 5G, the pilot signals that can be used for this purpose are called Sounding Reference Signals (SRS). The base station uses the uplink pilots from multiple users to select the downlink beamforming. This is the option that resembles what the textbooks on Massive MIMO are describing as the canonical form of the technology.
The second option is to let the base station transmit a set of downlink signals using different beams. The user device then reports back some measurement values describing how good the different downlink beams were. In 5G, the corresponding downlink signals are called Channel State Information Reference Signal (CSI-RS). The base station uses the feedback to select the downlink beamforming. The drawback of this approach is that 64 downlink signals must be transmitted to explore all 64 dimensions, so one might have to neglect many dimensions to limit the signaling overhead. Moreover, the resolution of the feedback from the users is limited.
In practice, the CSI-RS operation might be easier to implement, but the lower resolution in the beamforming selection will increase the interference between the users and ultimately limit how many users and layers per user that can be spatially multiplexed to increase the throughput.
New field trial based on SRS
The Signal Research Group has carried out a new field trial in Plano, Texas. The unique thing is that they confirm that the SRS operation was used. They utilized hardware and software from Ericsson, Accuver Americas, Rohde & Schwarz, and Gemtek. A 100 MHz channel bandwidth in the 3.5 GHz band was considered, the downlink power was 120 W, and a peak throughput of 5.45 Gbps was achieved. 8 user devices received two layers each, thus, the equipment performed spatial multiplexing of 16 layers. The setup was a suburban outdoor cell with inter-cell interference and a one-kilometer range. The average throughput per device was around 650 Mbps and was not much affected when the number of users increased from one to eight, which demonstrates that the beamforming could effectively deal with the interference.
It is great to see that “real” reciprocity-based Massive MIMO provides such great performance in practice. In the report that describes the measurements, the Signal Research Group states that not all 5G devices support the SRS-based mode. They had to look for the right equipment to carry out the experiments. Moreover, they point out that:
“Operators with mid-band 5G NR spectrum (2.5 GHz and higher) will start deploying MU-MIMO, based on CSI-RS, later this year to increase spectral efficiency as their networks become loaded. The SRS variant of MU-MIMO will follow in the next six to twelve months, depending on market requirements and vendor support.“
The following video describes the measurements in further detail:
Hi, Professor Bjornson.
Thanks for your informative post.
Three questions arise to my mind after reading the post.
1) You say that CSI-RS can be implemented easier than SRS in practice. Can you explain why it is so?
2) Can we implement FDD massive MIMO based on CSI-RS?
3) Imagine we do such an experiment with the mentioned test bed based on CSI-RS. Will the result be better or worse based on per device throughput?
1) This is the classical way of acquiring CSI. Since more signaling is needed, there are more details in the standard on how to implement it. There is no need for reciprocity calibration.
2) Yes, that is the benefit CSI-RS. It doesn’t require the uplink and downlink to be in the same band. One could even put the uplink on 4G and the downlink on 5G, in a dynamic spectrum sharing setup.
3) I guess that the previous experiments that I have reported on (from June this year and earlier) are based on CSI-RS. What you will see is that the reported throughput per layer might be the same, but they only send 8 layers, not 16 layers. This is where the real bottleneck appears. As you try to squeeze in more beams, you need a higher accuracy in the CSI to control the interference.
Hi Emil,
In your point of view what are the reasons for lower CSI-RS feedback?
I see the CSI-RS feedback will be based on WB or subband. However, even for the latter option the best scenario (subband) would be 4,8 or 16 (up to the total BW) RB granularity.
And for SRS a similar RB granularity (but also with higher RB granularity), however is not restricted to the total BW as CSI-RS. And those resources could also be multiplexed between users.
What I primarily meant with “the resolution of the feedback from the users is limited” is that we make measurements at the user side of high resolution; say 10 bits per sample. When we feed that back to the base station, we need to compress it using a codebook to just a few bits of resolution. This compression step reduces the ability to perform zero-forcing and similar interference suppression techniques in the spatial domain.
You are right that we can expect pilots to be transmitted on every subband (in SRS or CSI-RS). That is not a critical issue. If you have 1000 subcarriers, the channel will be different on each one of them, but the underlying channel might have only 100 taps in the time domain. So it would be enough to transmit pilots on every 10th subcarriers, estimate the channel in the time-domain and then reconstruct the channel on every subcarrier. Alternatively, one can design a good precoder at every 10th subcarrier and then interpolate between them: http://liu.diva-portal.org/smash/get/diva2:1048138/FULLTEXT01.pdf
By the way, at page 11 they mention: “today’s MU-MIMO trials are based on CSI-RS and are generally limited to no more than 8 layers”.
Do you see this as a rule-of-thumb?
Or just a lower likelihood to obtain more than 8 layers as CSI-RS has lower resolution, although on simulation shows that would be possible in certain conditions
I think the likelihood of benefiting in throughput from transmitting more than 8 layers is low, since the CSI quality is the limiting factor. The physical propagation environment is the same irrespective of the CSI acquisition protocol.
In simulations where the physical channel matches very well with the grid of beams used for CSI-RS, we will likely see that more than 8 layers can be used effectively. But apparently the people who make experiments claim that the gains disappear in practice.
Very nice!
In the video, the gentleman shows a single device with ~5.4Gbps data rate. I am just curious how it is possible, since as already it is mentioned in this blog post and in the video these data rate is overall rate for 8 devices not single device!
The software that is shown at 1:40 in the video has an Ericsson logo in the corner. I guess it shows how much data that the base station is transmitting, based on the choice of modulation and coding.
Thanks for a very insightful article. Indeed the true potential of massive MIMO will only be fully realized if the implementation moves away from CSI-RS with extremely limited feedback. This has not been happening yet, unfortunately.
TDD with “real” reciprocity is ideal to meet this challenge. Reciprocity calibration is a crucial part of achieving real reciprocity. 5G SRS, however, does not seem to include means of reciprocity calibration. I am wondering if you may have insight on this?
No, I don’t know how that is implemented in compliance with the NR standard. There are several types of algorithms in the academic literature. Some are based on transmitting a signal from one antenna port and measuring it at another port. Other algorithms are based on signals from the user. The LuMaMi testbed at Lund University has implemented some algorithms and concluded that the calibration errors are changing very slowly, so you might not need so much signaling to carry out the calibration, just track the changes slowly over time.