Massive MIMO: Where the Capacity Gain Actually Comes From

Ask what massive MIMO does and the usual answer is "beamforming." That's true and it undersells it considerably.

Beamforming improves the link to one user. The reason operators spend serious money on 64T64R radios is different: serving several users on the same time-frequency resources simultaneously, separated only by where they are in space. That's MU-MIMO, and it's where the capacity multiplier lives.

---

Three things antenna arrays buy you

Multiple antennas can be spent on three different objectives, and confusing them is the source of most massive MIMO confusion.

Diversity sends the same data over multiple paths so that a fade on one doesn't kill the link. Improves reliability, not throughput.

Beamforming concentrates energy directionally. Improves SNR and therefore link quality — which raises the achievable MCS and indirectly the throughput of one user.

Spatial multiplexing sends different data streams over paths the receiver can separate. This multiplies throughput directly.

Massive MIMO does all three at once, but the economics come from the third — applied across users rather than within one.

---

SU-MIMO vs MU-MIMO

SU-MIMO sends multiple streams to a single UE. The limit is the UE's own antenna count: a 4-antenna phone can receive at most 4 layers. NR permits up to 8 downlink layers per UE, but few devices support more than 4.

There's a harder practical limit too. Spatial multiplexing to one device needs a rich scattering environment so the streams arrive via distinguishable paths. In line-of-sight conditions with a compact handset, the channel matrix is poorly conditioned and rank collapses — you get one usable layer regardless of how many antennas either end has.

MU-MIMO sends streams to different UEs on the same resources. Now the separation comes from users being in different physical locations, which is far more reliable than hoping for multipath richness.

This is why massive MIMO is a cell capacity technology rather than a peak-rate technology. A single user doesn't see 5× throughput. The cell serves 5× the aggregate traffic in the same spectrum — which is what actually matters when you're dimensioning a network.

---

What "massive" means

The term implies the number of antenna elements is much larger than the number of simultaneously served users. That ratio is what makes the mathematics work.

Two useful properties emerge as arrays grow:

Channel hardening. As element count increases, the effective channel gain per user becomes more deterministic. Fast fading averages out across the array, so link adaptation gets more reliable and scheduling gets easier — you're no longer chasing a channel that changes every few milliseconds.

Favourable propagation. With enough elements, the channel vectors of spatially separated users become nearly orthogonal. Their signals stop interfering, and simple linear precoding gets close to optimal performance. You don't need exotic algorithms; you need enough antennas.

Decoding 64T64R

The notation confuses people because it doesn't say what most assume.

64T64R means 64 transmit and 64 receive chains — not 64 antenna elements. A typical 64T64R panel has 192 physical elements arranged in a grid, with groups of elements combined into 64 logical ports.

Common configurations: 32T32R and 64T64R for mid-band macro cells, 8T8R for low band, and 16T16R for lighter deployments.

The TRX count is what sets capability — it bounds how many independent beams the radio can form, and therefore how many users it can pair in MU-MIMO. It's also what sets cost, power draw, and weight, since each chain needs its own converters and amplifier. A 64T64R mid-band radio is a genuinely heavy object, and mast loading is a real deployment constraint.

---

Channel knowledge is the bottleneck

None of this works without knowing the channel. To form a beam toward a user — or to null interference toward another — the gNB needs the channel response for each element.

This is where TDD and FDD diverge sharply, and it explains a lot about how 5G was deployed.

TDD and reciprocity

In TDD, uplink and downlink use the same frequency. The channel is therefore essentially the same in both directions within the coherence time.

So the gNB measures the uplink using SRS (Sounding Reference Signals) and infers the downlink channel directly. It gets full channel knowledge for all 64 ports from one uplink transmission, and the feedback overhead is essentially independent of array size.

That last point is decisive. Reciprocity is why massive MIMO scales in TDD and doesn't in FDD — and why virtually all 5G mid-band deployment is TDD.

The catch: reciprocity holds for the propagation channel, not the hardware. Transmit and receive chains have different responses, so antenna calibration is required and must be maintained. Calibration drift degrades beamforming in ways that are genuinely hard to diagnose from network counters.

FDD and codebooks

In FDD, uplink and downlink are different frequencies with independent channels. Reciprocity doesn't apply.

The UE must measure downlink CSI-RS and report back using a codebook — a quantised set of precoding matrices. NR defines Type I (lower overhead, coarse) and Type II (higher resolution, designed for MU-MIMO, much heavier feedback).

The problem is that feedback scales with port count. A 32-port Type II report is large, and it's periodic. At 64 ports the overhead becomes prohibitive, which is why FDD massive MIMO is largely limited to 8 or 16 ports.

This is also why CSI feedback compression is one of the most active AI-in-RAN topics, and one of the few places 3GPP has taken machine learning into the air interface itself.

---

How the scheduler uses it

MU-MIMO turns scheduling into a pairing problem. Each slot, the scheduler must decide not just who to serve but who can be served together.

Pairing requires channels that are sufficiently orthogonal. Two UEs in nearly the same direction can't be separated — serving them together means each sees the other as interference and both get worse throughput than if they'd been served in sequence.

So the scheduler evaluates candidate groups against channel correlation, then picks precoding weights that maximise signal toward each paired user while nulling toward the others. This runs every slot, across hundreds of active UEs, with a hard deadline.

Real constraints that limit it:

SRS capacity. Uplink sounding resources are finite. In a busy cell, not every UE can be sounded frequently enough to support MU-MIMO pairing — which caps how many users are actually pairable regardless of array size.

Mobility. Channel estimates go stale. A fast-moving UE's channel changes within the coherence time, so precoding computed on old measurements points slightly wrong. MU-MIMO gains fall off with speed.

Uplink is harder. UEs have 1–2 transmit antennas and limited power. Uplink MU-MIMO exists but the gains are much smaller than downlink.

---

What you actually get

Realistic cell capacity gains from mid-band massive MIMO versus a 4T4R baseline land around 3–5× in dense urban deployments, less in rural cells where fewer users are simultaneously active. It's a load-dependent gain: an empty cell sees none of it, because there's nobody to pair.

Coverage improves separately, through beamforming gain — a 64-element array gives roughly 18 dB, meaningfully extending mid-band reach.

The costs are real. Power consumption is substantially higher, which is why network energy-saving features that shut down TRX chains at low load exist and matter. Weight and wind loading constrain which masts can carry the radio. Baseband processing scales with port count and user pairing complexity. And unit cost is several times a conventional radio.

The deployment logic follows from the load dependence: massive MIMO goes where users are dense. Urban macro cells and high-traffic venues justify it. Rural coverage cells generally don't.

---

The mental model

Beamforming improves one link. Spatial multiplexing multiplies capacity. MU-MIMO applies it across users.

Massive MIMO is a cell-capacity technology, not a peak-rate one. Gains scale with load.

64T64R = 64 chains, ~192 elements. TRX count sets capability, cost, and weight.

TDD reciprocity is the enabler. SRS gives full channel knowledge with overhead independent of array size.

FDD is codebook-limited, which is why FDD massive MIMO stalls around 8–16 ports.

SRS capacity and mobility cap real-world MU-MIMO pairing more than antenna count does.

If you keep one line: massive MIMO's value is reusing the same spectrum across spatially separated users — and everything about deployment, from TDD choice to where the radios go, follows from that.

---

Further reading

• Beamforming in 5G — architectures and beam management
• 3GPP TS 38.214 — Physical layer procedures for data, CSI reporting and codebooks
• 3GPP TS 38.211 — Physical channels, CSI-RS and SRS structure
• 3GPP TR 38.802 — Study on new radio access technology, physical layer aspects