Ferkans — Interactive Telecom Tutor

The Hardware Wall: Why ADC Resolution Limits Massive MIMO

The last ten chapters of this book have treated the receiver as a perfect linear operator: the channel output $\ntn{Y}$ arrives, and we decode. The premise breaks quietly but decisively at the massive MIMO scale. A contemporary mmWave base station is built around $n_r \in \{64, 128, 256\}$ receive antennas; each antenna feeds one (or two, for I/Q) analog-to-digital converter (ADC) that must sample at the full signal bandwidth. For 5G NR FR2, that bandwidth is 400 MHz–1 GHz; for 6G sub-terahertz proposals, it is 5–10 GHz. A high-resolution 12-bit ADC at 1 GSample/sec consumes roughly $1\,$ W per chain; at $n_r = 128$ this is $128\,$ W — before any amplifier, LNA, or digital signal processing.

The industry answer, adopted in nearly every serious massive-MIMO design since 2015, is low-resolution ADCs: $b \in \{1, 2, 3, 4\}$ bits per sample. Each bit saved roughly halves the ADC power. A 3-bit ADC at 1 GS/s is $\sim 10\,$ mW; a 1-bit ADC (a single comparator) is $\sim 1\,$ mW. Dropping from 12 to 3 bits saves two orders of magnitude in ADC power.

This is a radical redesign of the receiver. Every assumption about linear receivers — optimal MMSE filters, Gaussian noise at the decoder input, BICM demapper accuracy — was derived under the implicit premise that the ADC is essentially transparent. At $b \le 4$ the ADC becomes a dominant, strongly nonlinear distortion source, and the coded modulation must be redesigned around it: constellation shape, detection algorithm, and LDPC code design all shift.

We devote §§1-2 to the AWGN and 1-bit-MIMO consequences, and §5 to the full system picture. The take-home from §1 is quantitative: a $b$ -bit ADC contributes a quantisation noise floor that limits the effective SNR to $\mathrm{SNR}_q = 6.02\, b + 1.76\,$ dB, and for $b \le 3$ this number dominates the thermal noise at the amplifier output — meaning the code must be designed against a low-effective-SNR regime.

,

Definition:
Uniform Scalar Quantiser

A $b$ -bit uniform scalar quantiser is a deterministic map $Q: \mathbb{R} \to \mathcal{Q}$ , where $\mathcal{Q} \subset \mathbb{R}$ is a finite set of $2^b$ levels, spaced uniformly by step size $\Delta$ . If the input dynamic range is $[-A, A]$ then $\Delta = 2A / 2^b$ and the quantiser rounds its input to the nearest level: $Q(x) \;=\; \Delta \cdot \left\lfloor \frac{x}{\Delta} + \frac{1}{2} \right\rfloor, \qquad x \in [-A, A].$ For a complex input the quantiser operates independently on the real and imaginary parts: $Q(x) = Q(\mathrm{Re}\, x) + j\, Q(\mathrm{Im}\, x)$ .

The quantisation error is $e(x) = Q(x) - x$ . In the high- resolution regime $b \gg 1$ with input $x$ having a smooth density relative to $\Delta$ , Bennett's theorem (1948) asserts that $e$ is approximately uniform on $[-\Delta/2, \Delta/2]$ and statistically independent of $x$ , with variance $\mathbb{E}[e^2] = \Delta^2/12$ . This is the additive quantisation noise (AQN) model used in most linear system analyses. At low $b$ and low input SNR, the AQN model fails — the error becomes correlated with the signal and the receiver must treat quantisation as a genuinely nonlinear operator.

Distinguish carefully the quantiser $Q(\cdot)$ from the Q-function $Q(\cdot)$ used for AWGN error probabilities. They are different objects: $Q$ is a deterministic nonlinear map from continuous to discrete; $Q$ is an integral of a Gaussian density. Both appear in this chapter — e.g., the 1-bit capacity of §2 is expressed using $Q(\cdot)$ but refers to a channel whose front end is a $Q(\cdot)$ with 1 bit.

,

Definition:
Quantisation Signal-to-Noise Ratio ( $\mathrm{SNR}_q$ )

The quantisation SNR of a $b$ -bit uniform quantiser on a full-scale input $x \in [-A, A]$ is the ratio of signal power to quantisation-noise power: $\mathrm{SNR}_q \;=\; \frac{\mathbb{E}[x^2]}{\mathbb{E}[e^2]} \;=\; \frac{\mathbb{E}[x^2]}{\Delta^2 / 12}.$ For a full-amplitude sinusoidal input with $\mathbb{E}[x^2] = A^2/2$ and step $\Delta = 2A/2^b$ , this evaluates to $\mathrm{SNR}_q \;=\; \frac{A^2/2}{(2A/2^b)^2/12} \;=\; \frac{3 \cdot 2^{2b}}{2} \;=\; \frac{3}{2} \cdot 4^b,$ which in decibels is $\mathrm{SNR}_q^{(\mathrm{dB})} \;=\; 10 \log_{10}(1.5) + 20 b \log_{10} 2 \;\approx\; 1.76 + 6.02\, b \quad \mathrm{dB}.$ Each additional ADC bit buys $6.02\,$ dB of effective SNR: this is the per-bit rule of uniform quantisation, proved as the $\mathrm{SNR}_q$ theorem below.

The $6.02\, b + 1.76$ rule assumes (i) a uniformly-distributed or full-amplitude sinusoidal input and (ii) the input occupies the full ADC dynamic range. For a Gaussian input with $3\sigma$ clipping, the constant $1.76$ is replaced by $\approx -1.25\,$ dB (because 0.3% of samples clip and a Gaussian has higher peak-to- average ratio than a sinusoid). For small $b \le 3$ with a matched input, the real constant is closer to $-1$ to $+1\,$ dB — the approximation is tight but not exact. The BICM receiver rarely cares about the constant; it is the slope $6.02\, b$ that matters for system design.

,

Theorem: Quantisation-SNR Rule for a $b$ -Bit Uniform Scalar Quantiser

Under Bennett's high-resolution assumption (the input density is smooth relative to the quantisation step $\Delta$ ), the quantisation error $e = Q(x) - x$ is approximately uniform on $[-\Delta/2, \Delta/2]$ with variance $\mathbb{E}[e^2] = \Delta^2 / 12$ , and is uncorrelated with the input $x$ . For a $b$ -bit uniform quantiser with a full-scale input $x \in [-A, A]$ , $\mathrm{SNR}_q \;=\; \frac{\mathbb{E}[x^2]}{\mathbb{E}[e^2]}, \qquad \mathrm{SNR}_q^{(\mathrm{dB})} \;=\; 6.02\, b + 1.76 \;\; \mathrm{dB},$ for a full-scale sinusoidal input with $\mathbb{E}[x^2] = A^2/2$ .

The doubling of levels per extra bit means the quantisation step shrinks by a factor of 2, and the quantisation error power shrinks by 4 — that is, 6 dB. The $1.76\,$ dB constant is the specific signal-to-noise advantage of a full-scale sinusoid (peak-to- average-power ratio $2$ ) relative to a full-amplitude uniform distribution; it moves slightly for Gaussian inputs. What matters in system design is the $6\,$ dB-per-bit slope, which is robust across input distributions.

Show Hint

Compute $\Delta = 2A/2^b$ .

Use $\mathbb{E}[e^2] = \Delta^2/12$ for uniform $e$ .

Compute $\mathbb{E}[x^2] = A^2/2$ for a sinusoid.

Proof

Step 1 — Compute quantisation error variance

By Bennett's assumption, $e \sim \mathrm{Unif}(-\Delta/2, \Delta/2)$ independent of $x$ . Therefore $\mathbb{E}[e^2] \;=\; \int_{-\Delta/2}^{\Delta/2} \frac{t^2}{\Delta}\, dt \;=\; \frac{\Delta^2}{12}.$

Step 2 — Evaluate $\Delta$ for a $b$-bit ADC

With $2^b$ levels covering the range $[-A, A]$ , the step size is $\Delta = 2A / 2^b$ . Thus $\mathbb{E}[e^2] \;=\; \frac{(2A/2^b)^2}{12} \;=\; \frac{A^2}{3 \cdot 4^b}.$

Step 3 — Evaluate signal power

For a full-scale sinusoid $x(t) = A \cos(\omega t)$ , the time- averaged power is $\mathbb{E}[x^2] \;=\; \frac{A^2}{2}.$

Step 4 — Compute SNR and convert to decibels

$\mathrm{SNR}_q \;=\; \frac{A^2/2}{A^2/(3 \cdot 4^b)} \;=\; \frac{3 \cdot 4^b}{2} \;=\; 1.5 \cdot 4^b.KATEXPLACEHOLDER0END\mathrm{SNR}_q^{(\mathrm{dB})} \;=\; 10\log_{10}(1.5) + 20\, b\, \log_{10} 2 \;\approx\; 1.76 + 6.02\, b \;\; \mathrm{dB}. \qquad \blacksquare$ $

,

Achievable Rate vs. ADC Resolution at Fixed Input SNR

Achievable rate (bits per channel use) of a scalar AWGN channel observed through a $b$ -bit uniform ADC, as the ADC resolution $b$ varies from $1$ to $8$ bits. The channel input SNR is user-controllable. The plot overlays three curves: (1) the unquantised Shannon bound $\log_2(1 + \text{SNR})$ ; (2) the AQN-model approximation $\log_2(1 + \text{SNR}_{\mathrm{eff}})$ with $\text{SNR}_{\mathrm{eff}}^{-1} = \text{SNR}^{-1} + \mathrm{SNR}_q^{-1}$ ; (3) the exact quantised-AWGN mutual information (computed numerically by discretising the input). The AQN model is visibly tight at $b \ge 4$ and high input SNR; at $b \le 3$ or SNR $\le 0\,$ dB it underestimates the rate by a small margin. The curve plateaus at $\log_2(2^b) = b$ bits at high SNR — you cannot send more than $b$ bits through a $b$ -bit ADC.

Parameters

SNR [dB]10

Example: Effective Rate at a $3$ -Bit ADC with $\text{SNR} = 20$ dB

Consider a single receive antenna, Gaussian-input AWGN channel with pre-ADC SNR $\text{SNR} = 20\,$ dB, observed through a $3$ -bit uniform ADC. Under the additive-quantisation-noise (AQN) model: (a) Compute the quantisation-SNR in dB. (b) Compute the effective post-ADC SNR, $\text{SNR}_{\mathrm{eff}}$ . (c) Compute the achievable rate under AQN and compare it to the unquantised capacity. (d) Interpret: how much capacity is lost to quantisation?

Solution

(a) Quantisation SNR

From Thm. $b$ $b$ -Bit Uniform Scalar Quantiser" data-ref-type="theorem">TQuantisation-SNR Rule for a $b$ -Bit Uniform Scalar Quantiser: $\mathrm{SNR}_q = 6.02 \cdot 3 + 1.76 = 19.82\,$ dB.

(b) Effective post-ADC SNR

Converting to linear: $\text{SNR}_{\mathrm{lin}} = 100$ , $\mathrm{SNR}_{q,\mathrm{lin}} = 10^{19.82/10} \approx 95.9$ . By the AQN model (quantisation noise adds to channel noise), $\text{SNR}_{\mathrm{eff}}^{-1} \;=\; \text{SNR}^{-1} + \mathrm{SNR}_q^{-1} \;=\; 0.01 + 0.01043 \;=\; 0.02043,$ so $\text{SNR}_{\mathrm{eff}} \approx 48.9$ , i.e., $16.9\,$ dB.

(c) Rates

Unquantised capacity: $\log_2(1 + 100) = 6.66$ bits/ch.use. Quantised capacity (AQN): $\log_2(1 + 48.9) = 5.64$ bits/ch.use.

(d) Interpretation

The $3$ -bit ADC costs about $1.0$ bit/ch.use, or a $\sim 3\,$ dB equivalent SNR loss. This is substantial but tolerable: the saved $\sim 100\times$ in ADC power dwarfs the $1.0$ -bit rate loss for most massive-MIMO deployments. In practice the code is redesigned (stronger LDPC) to close part of this gap. $\blacksquare$

Historical Note: The Origin of Quantisation Theory: Bennett (1948) and Max (1960)

1948–1960

The statistical theory of quantisation was born at Bell Labs in 1948, the same year as Shannon's foundational information theory paper. W. R. Bennett analysed what we now call the additive quantisation noise model, showing that in the high-resolution regime a uniform quantiser's error is approximately uniformly distributed on $[-\Delta/2, \Delta/2]$ and independent of the input. Bennett's theorem is the theoretical foundation of the $6.02\, b + 1.76\,$ dB rule and remains in every DSP textbook.

The second milestone came in 1960, when J. Max (later at AT&T Bell Labs) formulated and solved the minimum-distortion quantiser problem: for a given input distribution and number of levels, find the quantiser levels and thresholds that minimise the mean-square error. The Lloyd-Max iteration — introduced concurrently by Lloyd and Max — is still the standard algorithm for non-uniform quantiser design. For a Gaussian input with $b$ bits, the Lloyd-Max quantiser is about $2.73\,$ dB better than the uniform quantiser at $b = 1$ (not enough to matter) and within $0.3\,$ dB at $b \ge 4$ (also not enough to matter). This is why uniform quantisers dominate in practice.

The modern resurgence of low-resolution quantisation theory (1996–present) was driven by digital communications: Widrow, Kollár, and Liu's 1996 paper revitalised the subject in the context of wireless receivers, and the massive-MIMO era (2015–) made 1–4 bit ADCs a mainstream engineering question.

,

Common Mistake: The AQN Model Fails at Low $b$ and Low SNR

Mistake:

A reader might take the additive quantisation noise (AQN) model — $y = Q(x) \approx x + e$ with $e$ uniform and independent of $x$ — as universally valid, and apply it even at $b = 1$ bit ADCs on AWGN with input SNR $\le 0\,$ dB. The LDPC-code design and the constellation design would then be computed from the AQN-effective SNR, with the assumption that Gaussian demapping and BICM receivers remain optimal.

Correction:

The AQN model relies on Bennett's assumption that the input density is smooth relative to the quantisation step $\Delta$ . At $b = 1$ there are only two levels ( $\pm A$ ) so $\Delta = 2A$ — a coarser grid than any realistic signal density. The error is strongly correlated with the sign of the input, not independent of it. Similarly at low SNR the input is dominated by noise, which has a wider spread than any Gaussian signal component; the quantiser error becomes bimodal. In both regimes, the AQN model underestimates the true quantised-capacity by up to $1$ – $2\,$ dB and misleads the code designer. The remedy, covered in §2, is to use the exact BSC capacity formula for 1-bit ADCs and to compute the quantised mutual information numerically for $b \in \{1, 2, 3\}$ . Never apply the $6.02\, b + 1.76$ rule to quantitatively-accurate code design below $b = 4$ .

,

⚠️Engineering Note

ADC Power Budget in 5G NR FR2 and 6G Research Prototypes

The ADC power crisis is a measurable, quantifiable engineering constraint, not an abstract concern. A 12-bit ADC at 1 GSample/sec dissipates roughly $1\,$ W (Walden figure-of-merit: $\sim 1\,$ pJ per conversion-step × $2^{12} \cdot 10^9$ steps/sec); a 3-bit ADC at the same sampling rate is $\sim 8$ pJ/sample × $10^9 = 8\,$ mW. Multiplied across $n_r = 128$ antennas:

12-bit: $128\,$ W (plus I/Q doubling: $256\,$ W) — infeasible.
3-bit: $\sim 1.6\,$ W total — compatible with 5G NR base station.
1-bit: $\sim 0.5\,$ W total — the 6G extreme.

3GPP TS 38.104 does not specify ADC bit-depth directly, but the receiver EVM (error-vector magnitude) requirements implicitly constrain it: FR2 UE category requires EVM $\le 8$ %, which corresponds to $\mathrm{SNR}_q \ge 22\,$ dB — hence $b \ge 4$ for the UE. For base stations (gNB), no comparable spec is published, but commercial massive-MIMO gNB implementations are known to use $b \in \{5, 6, 7\}$ with per-antenna automatic gain control (AGC) to exploit the dynamic range efficiently. Research prototypes at USC (Heath group), NYU (Rappaport), and KTH (Larsson group) have demonstrated 1-bit and 2-bit uplink ADCs achieving 80%+ of full-resolution uplink rate in $n_r \ge 64$ configurations.

Practical Constraints

•
UE category FR2 (mmWave): EVM $\le 8$ %, implying effective ADC $b \ge 4$
•
gNB side: $b \in \{5, 6, 7\}$ in commercial deployments, $b \in \{1, 2, 3\}$ in research prototypes
•
$n_r = 128$ antennas × 2 (I/Q) × $b = 12$ : intractable $\sim 256\,$ W
•
$n_r = 128$ × 2 × $b = 3$ : $\sim 1.6\,$ W — compatible with massive MIMO

📋 Ref: 3GPP TS 38.104 (NR base station Rx specifications); 3GPP TS 38.101-2 (UE Rx for FR2)

,

Key Takeaway

A $b$ -bit uniform ADC contributes a quantisation noise floor at effective SNR $6.02\, b + 1.76\,$ dB. For $b \ge 4$ this floor is above typical thermal noise, and the AQN model is accurate; a low-resolution ADC ( $b \in \{1, 2, 3\}$ ) introduces a dominant, signal-correlated nonlinearity that breaks the AQN model and forces exact or per-constellation-point analysis. The $6\,$ dB-per-bit rule drives a sharp practical trade: each saved bit buys roughly $2\times$ in ADC power, and for a 128- antenna mmWave receiver, dropping from $12$ to $3$ bits saves $\sim 100\times$ in ADC power at the cost of about $1$ bit/ch.use of capacity — an engineering trade that has come to define the 5G-NR FR2 and emerging 6G receiver architecture.

Quick Check

A $b$ -bit uniform ADC is used to sample a sinusoidal input. Going from $b = 8$ to $b = 10$ bits yields roughly how much improvement in quantisation SNR?

2 dB

6 dB

12 dB

60 dB

Correction:

12 dB

Correct. Each additional bit adds $6.02\,$ dB of quantisation SNR; 2 bits add $\sim 12\,$ dB.

Quick Check

The additive quantisation noise (AQN) model treats quantisation as independent additive noise. When does it FAIL to be a good approximation?

At high-resolution ADCs ( $b \ge 8$ ) with high input SNR

At low-resolution ADCs ( $b \le 3$ ) or at low input SNR

Only when the input is discrete (e.g., QAM)

Only at quantisation rates above the Nyquist rate

Correction:

At low-resolution ADCs (

b \le 3

) or at low input SNR

Correct. At $b = 1$ the input density is not smooth relative to the step; at low SNR the noise dominates. Both regimes violate Bennett's smoothness assumption.

ADC (Analogue-to-Digital Converter)

A device that samples a continuous-time signal at rate $f_s$ and quantises each sample into one of $2^b$ discrete levels ( $b$ = resolution). Key non-idealities: quantisation noise, clipping, timing jitter. In massive MIMO the ADC is the dominant per-antenna power consumer, motivating low-resolution ( $b \le 4$ ) architectures.

Quantisation SNR

The ratio of signal power to quantisation-noise power for a $b$ -bit uniform ADC. For a full-scale sinusoidal input, $\mathrm{SNR}_q = 6.02\, b + 1.76\,$ dB (Bennett's theorem). The $6\,$ dB-per-bit slope is robust across reasonable input distributions; the $+1.76\,$ dB constant is input-distribution- specific.

Low-Resolution DACs and ADCs

The Hardware Wall: Why ADC Resolution Limits Massive MIMO

Definition: Uniform Scalar Quantiser

Definition: Quantisation Signal-to-Noise Ratio (SNRq\mathrm{SNR}_qSNRq​)

Theorem: Quantisation-SNR Rule for a bbb-Bit Uniform Scalar Quantiser

Step 1 — Compute quantisation error variance

Step 2 — Evaluate $\Delta$ for a $b$-bit ADC

Step 3 — Evaluate signal power

Step 4 — Compute SNR and convert to decibels

Achievable Rate vs. ADC Resolution at Fixed Input SNR

Parameters

Example: Effective Rate at a 333-Bit ADC with SNR=20\text{SNR} = 20SNR=20 dB

(a) Quantisation SNR

(b) Effective post-ADC SNR

(c) Rates

(d) Interpretation

Historical Note: The Origin of Quantisation Theory: Bennett (1948) and Max (1960)

Common Mistake: The AQN Model Fails at Low bbb and Low SNR

ADC Power Budget in 5G NR FR2 and 6G Research Prototypes

Key Takeaway

Quick Check

Quick Check

ADC (Analogue-to-Digital Converter)

Quantisation SNR

Definition:
Uniform Scalar Quantiser

Definition:
Quantisation Signal-to-Noise Ratio ( $\mathrm{SNR}_q$ )

Theorem: Quantisation-SNR Rule for a $b$ -Bit Uniform Scalar Quantiser

Example: Effective Rate at a $3$ -Bit ADC with $\text{SNR} = 20$ dB

Common Mistake: The AQN Model Fails at Low $b$ and Low SNR