MIMO, Antenna Ports & Codebooks in 5G NR

Introduction

MIMO (Multiple-Input Multiple-Output) is the family of multi-antenna techniques that lets NR reuse the same resource blocks for several parallel data streams, or steer energy precisely at one UE, or serve many UEs at once on identical time-frequency resources. It sits in the physical layer, shaping how a transport block becomes layers, precoded symbols, and antenna-port transmissions before it ever hits the air. The layer/codeword mapping and precoding live in TS 38.211; the procedures that choose rank, precoder and CSI feedback live in TS 38.214; and the RRC that configures all of it lives in TS 38.331.

You meet MIMO the moment a UE moves past initial access into scheduled data. Every PDSCH and PUSCH transmission carries a rank (a layer count), rides one or two codewords, and is demodulated from DM-RS antenna ports. In FR2 (mmWave) MIMO is not optional at all — without beamforming the link simply does not close — and in FR1 massive-MIMO arrays are what deliver the cell-capacity gains operators actually deploy for.

It matters because MIMO is the single largest source of NR's spectral-efficiency advantage over LTE. Spectrum is fixed and expensive; antennas and the spatial dimension are comparatively cheap. Almost every headline throughput figure in a 5G data sheet — "up to 8 layers", "64T64R", "gigabit downlink" — is a MIMO claim, and understanding the layer/port/codebook machinery is what separates reading those numbers from knowing when they are real.

Why MIMO is needed

The concrete problem MIMO solves is that spectrum is scarce and Shannon caps how many bits a single channel can carry at a given SINR. You can widen the band or raise the power, but both run out fast. Space is the untapped axis: with several antennas at each end, a scattering channel presents multiple independent propagation paths, and each path can carry its own stream. Throughput then scales with the number of spatial streams rather than with bandwidth alone — the one lever that keeps giving.

Three things multi-antenna can buy you

People say "MIMO" as if it were one technique, but a multi-antenna array can be spent in three quite different ways, and NR uses all three, often at once. Knowing which one you are talking about is the first thing an interviewer probes.

These are not mutually exclusive. A typical FR2 transmission beamforms to reach the UE at all, then spatially multiplexes two or three layers inside that beam, while a cell-edge UE might drop to a single robust layer that behaves much like diversity. The rank the channel can support decides how far you can push multiplexing; everything below is machinery for choosing that rank and shaping the beams that carry it.

Layers, codewords, and antenna ports

Three terms sit at the heart of every MIMO discussion, and they are constantly confused. A codeword is a transport block after channel coding — one CRC, one coding chain, one MCS. A layer is one independent spatial stream of modulation symbols. An antenna port is a logical transmission entity defined by its reference signal. Data flows codeword → layers → ports → physical antennas, and each arrow is a defined mapping in TS 38.211.

The rank is simply the number of layers in a transmission, and it is the single most important MIMO parameter. Rank 1 is one robust stream; higher rank is more parallel streams for strong, well-conditioned channels. NR supports up to 8 layers in the downlink per UE and up to 4 layers in the uplink per UE (TS 38.211, TS 38.214).

Why two codewords at all? Coding and mapping a very large transport block onto up to eight layers as a single codeword would make the block enormous and the HARQ retransmission of a whole block wasteful. So NR splits into a second codeword once the rank climbs past four, letting each codeword carry its own MCS and its own HARQ acknowledgement. The split point is fixed: one codeword for ranks 1–4, two codewords for ranks 5–8.

Now the antenna port. This is the idea people get wrong most often: an antenna port is not a physical antenna. It is a logical channel identity, defined so that "the channel over which one symbol on the port is conveyed can be inferred from the channel over which another symbol on the same port is conveyed." Concretely, each port carries a known reference signal — a specific DM-RS pattern for data — and the receiver estimates that port's effective channel from those pilots. Whatever precoding or beamforming the transmitter applies behind the port is invisible to the receiver; it only ever sees an effective channel per port.

The port numbers are not arbitrary either: TS 38.211 assigns fixed ranges per signal, so a port number tells you what you are looking at. PDSCH DM-RS uses ports numbered from 1000, PUSCH DM-RS from 0, CSI-RS from 3000, and SRS from 1000 in the uplink. A downlink rank-4 PDSCH therefore occupies DM-RS ports 1000–1003, one per layer.

The full chain for a 2-layer downlink is shown below: one codeword is scrambled and modulated, mapped to two layers, multiplied by a precoding matrix W, emitted on two antenna ports each carrying its own DM-RS, and finally radiated by the physical array. The channel mixes everything in the air; the receiver uses the per-port pilots to invert the mix and recover the two layers.

SU-MIMO, MU-MIMO, and massive MIMO

Spatial multiplexing can serve one user with many layers, or many users with a few layers each, on the very same resource blocks. That choice is the difference between SU-MIMO and MU-MIMO, and it changes what kind of channel feedback you need.

Massive MIMO is not a separate transmission scheme; it is a scale of array. Its value is that with dozens of antenna elements the gNB can form many highly directional beams and place a spatial null exactly where each co-scheduled UE's neighbour sits. That is what turns MU-MIMO from a theoretical gain into a deployed one: a 64-element array can comfortably co-schedule a handful of spatially separated users, each seeing near-interference-free layers, so the cell delivers several times the throughput of a single-user system on the same spectrum. The catch is feedback — the gNB can only null what it can accurately measure, which is exactly why NR introduced the high-resolution Type-II codebook discussed next.

Downlink precoding and the CSI codebooks

Between the layers and the antenna ports sits precoding: a matrix multiply that maps L layers onto P ports, shaping the transmitted beams so the streams arrive separable at the receiver. For the gNB to precode well it must know the downlink channel. In non-reciprocal deployments (typically FDD) it cannot measure the downlink itself, so the UE measures a CSI-RS and reports back a recommended precoder chosen from a shared codebook — a standardised dictionary of precoding matrices. NR (TS 38.214) defines two families with a deliberate trade-off between feedback size and precision.

Both families describe the antenna panel by a pair of oversampled DFT dimensions (N₁, N₂) — the number of cross-polarised antenna columns and rows — and oversampling factors (O₁, O₂), all carried in the RRC codebookConfig. A precoder is then a selection of DFT beams from that grid plus per-polarisation co-phasing. That is the shared skeleton; the two types differ in how richly they describe the beam(s).

Type-I is the workhorse. Its precoders are essentially DFT beams: the UE picks the best beam (plus a co-phasing term across polarizations) per layer. It comes in a single-panel form for ordinary arrays and a multi-panel form for arrays built from several sub-panels, and it can be reported as one wideband selection for the whole carrier or refined per subband when frequency-selective precoding helps. Feedback is compact, and for SU-MIMO that coarse "point at the strongest beam" behaviour is entirely adequate.

Type-II is the high-resolution codebook built for MU-MIMO. Instead of one beam per layer, it reports a weighted linear combination of several orthogonal DFT beams (typically L = 2, 3 or 4 beams), each with quantised amplitude and phase, per subband. That richer description lets the gNB reconstruct the channel accurately enough to steer energy at one user while placing a null at a co-scheduled user. The price is a much larger report, so Type-II is used selectively where the MU-MIMO capacity payoff justifies the overhead. Later releases added an enhanced Type-II (eType-II, Rel-16) that compresses this feedback in the frequency domain using a DFT basis, cutting overhead for higher ranks.

Which codebook, how many beams, the subband size, and the maximum rank are all set by RRC through the CSI-ReportConfig and its codebookConfig, tied to a CSI-ResourceConfig that points at the CSI-RS resources the UE should measure. The network thus dials the exact feedback fidelity it is willing to pay for, per report, per UE.

CSI feedback: RI, PMI, CQI — and reciprocity

The codebook is only half the loop. The UE ties its channel measurement into one actionable recommendation using three headline quantities, each measured against the CSI-RS and reported back to the gNB.

The order matters: the UE first decides how many layers are worth attempting (RI), then which precoder best serves that rank (PMI, a Type-I or Type-II selection), then what data rate that combination can sustain (CQI). A Layer Indicator (LI) can additionally flag the strongest layer for phase-tracking reference-signal placement. The gNB treats all of this as a recommendation it is free to override — it sees every UE's report and the overall load — but it is the starting point for choosing rank, precoder, and MCS.

The report is not always the same shape: RRC configures a reportQuantity (for example cri-RI-PMI-CQI, or cri-RSRP for beam management) and a timing behaviour that is periodic (on PUCCH), semi-persistent (PUCCH or PUSCH, activated by MAC CE), or aperiodic (on PUSCH, triggered by a CSI request field in an uplink DCI format 0_1). The CRI (CSI-RS Resource Indicator) picks which measured resource the rest of the report refers to when several are configured.

There is a shortcut, and it hinges on duplex mode. In TDD, uplink and downlink share the same frequency, so the channel is reciprocal: the gNB can estimate the downlink directly from an uplink SRS and compute its own precoder, skipping PMI feedback entirely. This is the natural home of massive MIMO, because reciprocity gives the gNB full-resolution channel knowledge for free. In FDD, uplink and downlink are on different frequencies and the channel is not reciprocal, so the UE must feed back an explicit PMI from the codebook. That is precisely why the high-resolution Type-II codebook matters most in FDD MU-MIMO: it is the only way to hand the gNB the channel precision that a TDD system would simply measure.

Uplink MIMO: codebook vs non-codebook

The uplink flips the geometry — the small, power-limited UE is now the transmitter and the gNB is the sophisticated receiver — so NR defines its own two precoding modes, selected by RRC via txConfig. Both rely on the SRS (Sounding Reference Signal), which the UE transmits so the gNB can measure the uplink channel.

The coherence subset is a uniquely uplink concern worth dwelling on. A fully coherent UE can control the relative phase across all its transmit antennas, so it can use any precoder in the codebook, including ones that combine antennas constructively. A non-coherent UE can only send one layer per antenna with no cross-antenna phase control, so its codebook shrinks to antenna-selection precoders. Partial coherence sits between. The gNB never signals a TPMI the UE cannot execute, so this capability directly shapes the achievable uplink precoding.

Uplink MIMO also caps at 4 layers and a single codeword, versus 8 layers and up to two codewords in the downlink. The following table lines up the two directions.

So the two reference signals divide the labour cleanly: CSI-RS is the downlink sounding signal the UE measures to feed back RI/PMI/CQI, while SRS is the uplink sounding signal the gNB measures to pick the uplink precoder (and, under reciprocity, the downlink one too). Get those two roles straight and most of NR's MIMO signalling falls into place.

DM-RS ports, CDM groups, and the TCI/QCL link

Two final pieces make beamformed MIMO actually decodable: the pilots that separate the layers, and the signalling that tells the UE which beam a transmission arrives on.

Each layer rides an antenna port whose channel the receiver estimates from that port's DM-RS. To pack several ports into the same resource elements without them colliding, NR groups ports into CDM groups (Code Division Multiplexing groups). Ports within one CDM group share the same resource elements but are separated by orthogonal cover codes across frequency and time; ports in different CDM groups sit on different subcarriers (a frequency comb). This is why the DCI field Antenna Port(s) also tells the UE the number of CDM groups without data — the UE must know which resource elements are occupied by other users' DM-RS (and therefore carry no data for it) to rate-match correctly. DM-RS comes in Type 1 (up to 4 ports per CDM group over 2 groups = 8 ports) and Type 2 (up to 6 ports per group over 3 groups = 12 ports), and in single- or double-symbol variants that trade pilot overhead for more ports.

Beamforming adds one more question the UE cannot answer on its own: which receive beam should I point at this transmission, and what channel properties can I assume it shares with a signal I have already measured? That is the job of the TCI (Transmission Configuration Indication) state and its QCL (Quasi-Co-Location) relationship. A TCI state tells the UE that the DM-RS of an incoming PDSCH or PDCCH is quasi-co-located with a specified source reference signal — usually an SSB or a CSI-RS — with respect to a set of channel properties.

QCL-TypeD is the one that makes analog beamforming work: it points the UE at the same spatial filter (receive beam) it used for the source reference signal, so it can physically aim its array before the data even arrives. TCI states are configured by RRC and the active one is confirmed by a MAC CE (or indicated in DCI for some channels), which is how the network switches a UE's serving beam as it moves. Without the TCI/QCL framework, a UE in FR2 would have no idea which of its narrow beams to listen on, and beamformed MIMO would simply not close.

Summary

MIMO spends a multi-antenna array three ways — spatial multiplexing for throughput, diversity for reliability, beamforming for coverage — and NR uses all three at once. The vocabulary that matters is the chain codeword → layer → antenna port → physical antenna: rank is the layer count (up to 8 DL, 4 UL), codewords split at rank 5, and a port is a logical identity defined by its DM-RS, never a physical element.

The whole machine is steered by channel knowledge. In FDD the UE feeds back RI/PMI/CQI from a codebook — coarse Type-I for SU-MIMO, high-resolution Type-II for FDD MU-MIMO; in TDD the gNB skips PMI and exploits reciprocity from the uplink SRS. Uplink adds its own codebook-vs-non-codebook choice and the coherence-subset constraint. Finally, DM-RS ports and CDM groups separate the layers at the receiver, while TCI/QCL-TypeD tell it which beam to point at — the piece without which FR2 beamformed MIMO simply would not close.

Quick Q&A

The questions interviewers reach for when they want to know whether you actually understand NR MIMO rather than just its vocabulary.