Massive MIMO - What is it ?

As the name implies, Massive MIMO is an antenna array system using Massive Amount of Antenna. It is also called Large Scale MIMO.

Then you may ask "How many Antenna is required to be called 'Massive (Large Scale) MIMO' ?".

The answer may be different depending on how you design the reciever algorithm (equalizer design), but assuming that we use the simplest (the most primitive) reciever design, we may need over 300 antenna. If you think this is too big number, you may decrease the number by applying a little bit smarter reciever (equalizer) algorithm.

In some of the prototype system that can be called as Massive MIMO, I see the number to be 64, 256 but you would see different numbers as more prototypes are implemented. According to R1-163132 section 2.5, it is said that the number of antenna will be up to 256 in DL and 32 in UL in 5G (Ref [7]).

In conventional LTE using a normal MIMO, the maximum number of antenna in MIMO as of now (Mar 2015) is 8 x 2 or 4 x 4 and recently (Apr 2016) even 8 x 8 is mentioned.

Two things about that number are worth saying early. The count that matters is not really the absolute number of antennas. It is the ratio between antennas at the base station and users served at the same moment. When that ratio is large, the channels to different users become close to orthogonal, and simple linear processing separates them.

The second thing is that antennas counted in a research paper and antennas counted in a specification are rarely the same objects. Antenna elements, transceiver units and antenna ports are three different things. Discussions of Massive MIMO move between them without warning, so the first section below separates them.

For further details of the Massive MIMO, refer to following pages. I strongly recommend to read the Why we need it page at least and leave other pages as an optional. Each entry below opens a short section on this page, and every section links out to the full page for that topic.

Antenna elements, TXRUs and antenna ports are three different counts

The 256 figure quoted above counts antenna elements. That is the physical count, the number of radiating patches on the panel. It is the number an antenna engineer works with, and it decides the size and the aperture of the array.

A transceiver unit, or TXRU, is a complete transmit and receive chain. It carries its own converter, mixer and amplifier. TXRUs are expensive, so a large array rarely has one per element. Several elements are instead wired to one TXRU behind a fixed phase shift, and each such subarray forms one fixed narrow beam.

An antenna port is neither of those. 3GPP defines a port by a reference signal rather than by hardware. Two symbols sit on the same port when the channel carrying one can be inferred from the channel carrying the other. A port is what the UE can measure, and nothing more.

The specification limits ports, not elements. In the Rel-17 ASN.1 the CSI-RS resource declares nrofPorts as ENUMERATED {p1, p2, p4, p8, p12, p16, p24, p32}. So 32 is the most a UE is ever asked to measure, whatever sits behind those ports.

< Figure 1. Elements, TXRUs and ports >

Figure 1 puts the three counts side by side. Elements are grouped onto TXRUs, and TXRUs are mapped to the ports that carry CSI-RS. Each step to the right reduces the count, and each step hides detail from the UE.

A panel with 256 elements can therefore present 32 ports, and the UE never sees the other 224. The array stays large while the measurement problem stays small. Much of the engineering in Massive MIMO lives in that mapping.

Elements are physical : They set the aperture and the size of the panel.
TXRUs are the cost : One per element is rare, so elements sit in subarrays behind fixed phase shifts.
Ports are what the UE measures : NR caps CSI-RS at 32 ports, however many elements are behind them.

Why we need it ?

Why we need it argues that Massive MIMO in 5G is less a choice than a consequence. High frequency is the starting point.

A millimetre wave antenna element is small, because element size follows wavelength. A small element captures little energy, so one element alone gives a poor link budget at those frequencies. The remedy is to use many elements and combine them coherently. The array recovers the aperture that the single element lost.

That gives the first benefit, which is array gain. Directing energy one way instead of spreading it in every direction gains roughly a factor of M with M elements. The second benefit is spatial separation, because narrow beams let one time and frequency resource serve several users at once.

I would read that page before any of the others. It is the one that explains why the industry accepted the cost and complexity of large arrays at all.

How it is implemented in 5G specification ?

Everything above is physics. How it is implemented in 5G specification is where it turns into signalling, and the answer is the CSI-RS codebook.

A gNB cannot form a useful beam without knowing the channel. In the downlink it transmits CSI-RS and the UE measures those reference signals. The UE then reports an index into a table both sides already hold. The table is the codebook, and the index is the PMI.

NR defines several codebook families. Type I is coarse and cheap, and it suits single user beam selection. Type II is far more detailed, since it reports a weighted combination of beams rather than one beam. Type II exists mainly so that MU-MIMO can null the interference between users.

Enhanced Type II then compresses that report, because a full Type II report is large enough to become a problem in itself. The governing spec text is 38.214-5.2.2.2, and the configuration sits in CodebookConfig in 38.331.

Channel Model

A channel model is what you simulate against before hardware exists. Channel Model collects the models used for Massive MIMO work.

No single model covers every situation, so the practice is to define a few typical cases and derive a model for each. That page gives point to point MIMO, multi user MIMO, Massive MIMO with small cells, and FD-MIMO in three dimensions.

Going from few antennas to many changes what a model has to get right. With a handful of antennas the interesting question is the fading statistics on each path. With hundreds, the interesting questions become how correlated neighbouring elements are, and whether two users are far enough apart in angle to be separated.

Channel estimation sits on the same page. That is where the modelling assumptions stop being free and start costing real reference signal resources.

Reciever Model

The definition at the top of this page says over 300 antennas with the most primitive receiver, and fewer with a smarter one. Reciever Model is where that trade is examined.

The primitive option is the linear receiver. Matched filter, zero forcing and MMSE all treat the array as a set of weights, and they cost roughly one matrix operation per symbol. The accurate option is maximum likelihood detection, which searches over transmitted vectors and becomes impossibly expensive as antennas increase.

The useful result is that the gap between the two narrows as the array grows. When antennas greatly outnumber users, the channel vectors become close to orthogonal, and even a matched filter separates users well.

That is a large part of why Massive MIMO is practical. The array is what makes the cheap receiver good enough. The antenna count buys back processing that a smaller system would have to spend.

FD-MIMO

FD stands for Full Dimension. FD-MIMO is the step from an array that steers only in the horizontal plane to one that steers in both planes.

A conventional base station antenna is a vertical column of elements with a fixed downtilt. It aims left and right, but not up and down. A two dimensional panel can aim at a particular floor of a building as well as a particular direction. That matters when users are stacked vertically.

This is where the N1 by N2 codebook shape gets its physical meaning. N1 counts one dimension of the port grid and N2 counts the other. A (16,1) layout is a purely horizontal array, while (4,4) is a square panel that steers in both planes. Both give 32 ports, and they behave nothing alike.

FD-MIMO began in LTE, in TR 36.897 and TR 36.873. The linked page covers form factor, transceiver architecture and how to model the antenna beam pattern.

MU-MIMO

MU-MIMO serves several UEs on one time and frequency resource, and separates them by space alone. The idea is not new. LTE has it in TM5 and WLAN has its own version, but the scale changes completely with a large array.

MU-MIMO is what the antenna to user ratio buys you. With a small array, two users close together in angle interfere badly and cannot share the resource. With a large array the beams are narrow enough that the same two users are separable.

The cost is precision. Nulling the interference between users needs accurate channel knowledge, and a coarse beam index is not enough. That is exactly why Type II codebooks exist, and why they report so much more than Type I.

The linked page covers the mathematical model, resource allocation and the effect on throughput.

Technical Challenges

Technical Challenges is the honest part of the topic. Massive MIMO is agreed as a core 5G technology, but agreeing on it did not finish it.

The open items are practical rather than theoretical. The linked page groups them roughly as follows.

How to arrange the antennas in the array.
How to model a three dimensional channel.
How to make it work in FDD, where reciprocity is not available.
How to generate a wide beam from an array built to make narrow ones.
How to keep hundreds of chains calibrated against each other.
How to handle the complexity of scheduling and precoding in real time.

Calibration deserves a note of its own. Every claim about coherent combining assumes the phase relationship between chains is known. Components drift with temperature and with age, so an uncalibrated array slowly stops forming the beam it believes it is forming.

Scheduling deserves another. Choosing which users to pair, on which resources, with which precoder, is a combinatorial problem. It has to be solved again every slot, which puts a hard limit on how clever the algorithm can afford to be.

Why does Massive MIMO prefer TDD ?

One question is missing from the list above, and the answer decides how much of this is usable in practice. How does the base station learn the channel to every one of its ports ?

In TDD both directions share a carrier, so the channel is reciprocal. Each UE sends one uplink pilot, and the base station measures that pilot on all of its antennas at once. The pilot cost grows with the number of users, and not with the number of antennas.

In FDD the two directions sit on different frequencies, and reciprocity no longer holds. The base station has to send a reference signal from every port, and the UE has to measure them and report back. The cost now grows with the number of ports, both in the downlink and in the feedback.

< Figure 2. Why TDD scales and FDD does not >

Figure 2 shows the two arrangements next to each other. The left side needs one arrow per user. The right side needs one arrow per port, plus a report coming back.

That asymmetry explains several things at once. It is why the research literature is largely TDD, and why FDD operation appears as an open problem under Technical Challenges. It also explains the port cap in the previous section.

The 32 port limit is not a limit on the hardware. It is a limit on what can be measured and reported at a sensible cost. NR does support large arrays in FDD, through codebooks instead of reciprocity. Enhanced Type II exists because the report needed for FDD MU-MIMO was otherwise too large to carry.

TDD scales with users : One uplink pilot per UE reveals the channel on every antenna.
FDD scales with ports : Every port needs its own reference signal and its own feedback.
The port cap follows from this : 32 is what the reporting budget allows, not what the panel allows.

Reference

[1] GFDM Interference Cancellation for Flexible Cognitive Radio PHY Design

R. Datta, N. Michailow, M. Lentmaier and G. Fettweis

Vodafone Chair Mobile Communications Systems,

Dresden University of Technology,

01069 Dresden, Germany

Email:[rohit.datta, nicola.michailow, michael.lentmaier, fettweis]@ifn.et.tu-dresden.de

[2] 5G NOW. D3.1 5G Waveform Candidate Selection

[3] Massive MIMO and Small Cells : Improving Energy Efficiency by Optimal Soft-Cell Coordination

Emil Bjornson, Marios Kountouris and Merouane Debbah

Alcatel-Lucent Chair on Flexible Radio, SUPELEC, Gif-sur-Yvette, France

Department of Telecommunications, SUPELEC, Gif-sur-Yvette, France

ACCESS Linnaeus Center, Signal Processing Lab, KTH Royal Institue of Technology, Stockholm, Sweden

[4] Massive MIMO Info Point

[5] Massive MIMO for Next Generation Wireless Systems

Erik G. Larson, ISY, Linkoping University, Sweden

Ove Edfors, Lund University, Sweden

Fredrik Tufvesson, Lund University, Sweden

Thomas L. Marzetta, Bell Labs, Alcatel-Lucent, USA

[6] Scaling up MIMO : Opportunities and Challenges with Very Large Arrays

Fredrik Rusek, Dept. of Electrical and Information Technology, Lund University, Lund, Sweden

Daniel Persson, Dept. of Electrical Engineering (ISY), Linkoping University, Sweden

Buon Kiong Lau, Dept. of Electrical and Information Technology, Lund University, Lund, Sweden

Erik G. Larsson, Dept. of Electrical Engineering (ISY), Linkoping University, Sweden

Thomas L. Marzetta, Bell Laboratories, Alcatel-Lucent, Murray Hill, NJ

Ove Edfors, Dept. of Electrical and Information Technology, Lund University, Lund, Sweden

Fredrik Tufvesson, Dept. of Electrical and Information Technology, Lund University, Lund, Sweden

[7] 3GPP R1-163132. 3GPP TSG RAN WG1 Meeting #84bis - Discussion on the frame structure design for NR

(click here for TDocs)