Ferkans — Interactive Telecom Tutor

The Real Metric is Bits per Joule

Sections 19.1–19.4 treat resolution $b$ as a design parameter and the achievable rate $R(b)$ as the objective. In deployment the more meaningful metric is energy efficiency

$\eta = \frac{R(b)}{P_{\text{rx}}(b)},$

where the denominator collects the power of the ADCs themselves, the baseband processing they drive, and the RF front-end they follow. The energy-rate tradeoff is not monotone in $b$ : at very low $b$ rate is cut too aggressively; at very high $b$ the power grows exponentially while rate saturates. There is an interior optimum whose position depends on SNR, array size, and the Walden figure-of-merit of the target ADC technology. Identifying this optimum is the job of this section.

,

Definition:
Receiver Energy Efficiency

The energy efficiency of a low-resolution massive-MIMO uplink receiver with uniform resolution $b$ per antenna is

$\eta(b) = \frac{W\,\sum_{k=1}^{K} R_k(b)} {N_r\,(P_{\text{ADC}}(b) + P_{\text{RF}}) + P_{\text{BB}}(b)} \quad \text{[bits/J]},$

where $W$ is the system bandwidth, $R_k(b)$ is the per-user rate (bits/channel use) derived via Bussgang analysis, $P_{\text{ADC}}(b) = c_0\,2^b\,f_s$ is the per-ADC power, $P_{\text{RF}}$ is a fixed RF front-end term, and $P_{\text{BB}}(b) \propto N_r^{2} \cdot b$ approximates the baseband processing power (MAC width grows with $b$ ). The dependence on $b$ is exponential in the ADC term and polynomial in the baseband term.

The numerator grows slowly (logarithmically) with $b$ because $\kappa_b \to 1$ quickly. The denominator grows multiplicatively in $b$ . Their ratio is unimodal, with an interior maximum between $b \approx 2$ and $b \approx 5$ depending on the operating point.

Theorem: Optimal Resolution at Low Per-Antenna SNR

Consider the single-user uplink with $N_r$ antennas and per-antenna SNR $\text{SNR}$ so small that $N_r\,\text{SNR} \ll 1$ . In this regime the per-antenna rate is approximately linear in $\kappa_b$ , so $R(b) \approx \kappa_b\,N_r\,\text{SNR}/\ln 2$ , and the energy efficiency (ignoring the baseband term) is

$\eta(b) \approx \frac{\kappa_b\,N_r\,\text{SNR}/\ln 2} {N_r\,c_0\,2^b f_s + \text{const}}.$

The maximizer is $b^\star = 1$ (the 1-bit receiver) whenever the ADC dominates the denominator. In words: at low per-antenna SNR the 1-bit receiver is energy-optimal because doubling $b$ roughly doubles denominator while only bringing $\kappa_b$ from $2/\pi$ to $0.88$ — a factor $1.4$ improvement in rate at the cost of a factor $2$ in power.

The $2/\pi$ fraction pins the 1-bit rate and the Walden law pins the cost. At low SNR the rate gap is small in absolute terms and the 1-bit energy advantage is decisive. Mixed-ADC becomes attractive only when the per-antenna SNR rises into the moderate regime where the remaining 1.96 dB gap starts to matter.

Proof

Low-SNR rate

$R(b) = \log_2(1 + \kappa_b\,N_r\,\text{SNR}) \approx \kappa_b\,N_r\,\text{SNR}/\ln 2$ for $\kappa_b\,N_r\,\text{SNR}\ll 1$ .

Numerator ratio

$\kappa_2/\kappa_1 = 0.883/0.637 \approx 1.386$ , $\kappa_3/\kappa_2 \approx 1.094$ , $\kappa_\infty/\kappa_b \to 1/\kappa_b$ . The diminishing returns are visible even in the ratio sequence.

Denominator growth

$P_{\text{ADC}}(b) = c_0 2^b f_s$ grows by a factor $2$ per bit. Hence $\eta(b+1)/\eta(b) = (\kappa_{b+1}/\kappa_b)\cdot(1/2)$ , which is less than 1 for $b \geq 1$ whenever $\kappa_{b+1} < 2\kappa_b$ — always true for $b \geq 1$ under the Gaussian-optimized quantizer table. So $b^\star = 1$ . $\blacksquare$

,

At Moderate SNR the Optimum Slides to 3–4 Bits

The story at moderate-to-high SNR is not the same. When $\kappa_b\,N_r\,\text{SNR}$ is comfortably above 1, the rate scales as $\log_2$ and further reductions in $\rho_b$ still pay off. Numerical studies (Björnson 2017 figures 5-6, Jacobsson 2017 figures 10-11) show that at $\text{SNR} = 5$ dB with $N_r$ in the hundreds the energy-efficient optimum moves to $b^\star \in \{3, 4\}$ . Further increasing $b$ past 4 rarely helps: the rate ceiling is approached but the ADC power doubles again for no rate in return.

A pragmatic design rule that fits most deployments is:

Low per-antenna SNR ( $\leq -5$ dB, many antennas): $b^\star = 1$ bit, optionally add a small mixed-ADC fraction for channel estimation.
Moderate SNR ( $-5$ to $10$ dB): $b^\star \in \{3, 4\}$ with uniform allocation.
High SNR ( $\geq 10$ dB): $b^\star \in \{5, 6\}$ ; low-resolution rarely helps and a conventional high-resolution receiver is preferred.

,

Energy Efficiency vs ADC Resolution

For fixed $N_r$ , $\text{SNR}$ , and Walden FoM, plot $\eta(b)$ in bits/joule as a function of $b \in \{1,\ldots,8\}$ . Observe the interior maximum. Vary the per-antenna SNR to see the optimum slide from $b^\star = 1$ (low SNR) to $b^\star \in \{4,5\}$ (high SNR).

Parameters

Number of RX antennas128

Per-antenna SNR (dB)-5

Bandwidth (MHz)200

Walden FoM (fJ / conv-step)100

Fixed RF power per antenna (mW)30

Example: Break-Even Between 1-Bit and 4-Bit

A base station has $N_r = 128$ antennas, $W = 200$ MHz, $f_s = 250$ MS/s, Walden FoM $= 100$ fJ/conversion-step, and a fixed per-antenna RF cost $P_{\text{RF}} = 30$ mW. Compare the 1-bit and 4-bit single-user uplink energy efficiencies at per-antenna SNRs of $-10$ dB and $0$ dB. Which resolution wins in each case?

Solution

Per-ADC powers

$P_{\text{ADC}}(1) = 100\cdot 10^{-15}\cdot 2 \cdot 250\cdot 10^6 = 50$ µW. $P_{\text{ADC}}(4) = 100\cdot 10^{-15}\cdot 16\cdot 250\cdot 10^6 = 400$ µW. Both are dwarfed by $P_{\text{RF}} = 30$ mW, which is the same for both cases.

Total denominator per antenna

$1$ -bit: $P_{\text{RF}} + P_{\text{ADC}}(1) \approx 30.05$ mW. $4$ -bit: $P_{\text{RF}} + P_{\text{ADC}}(4) \approx 30.40$ mW. Ratio $\approx 1.012$ — a negligible 1.2% power penalty for the extra precision.

Rates

At $-10$ dB: $\kappa_1 \cdot 128 \cdot 0.1 \approx 8.15$ , $R_1 \approx \log_2(1 + 8.15)\approx 3.19$ bit/use. $\kappa_4 \cdot 128 \cdot 0.1 \approx 12.68$ , $R_4 \approx \log_2(1 + 12.68) \approx 3.77$ bit/use. Rate ratio 1.18, power ratio 1.012, so EE ratio $\approx 1.17$ : 4-bit wins even at $-10$ dB because the RF term dominates and the rate gain is free.

Revisit with a low-RF scenario

Repeat with $P_{\text{RF}} = 0$ (hypothetical). Now denominators are $50$ and $400$ µW, ratio $8$ . Rate ratio still $1.18$ , EE ratio $1.18/8 \approx 0.15$ , so 1-bit wins by a factor of 7. The conclusion: the energy-efficiency argument for 1-bit strongly depends on how small $P_{\text{RF}}$ is relative to $P_{\text{ADC}}$ . For mmWave radios where both scale with bandwidth, the 1-bit advantage is most pronounced; at sub-6 GHz with a dominant RF term, the ADC is a minor player. $\blacksquare$

The Baseband Processor is Not Free

A complete energy accounting must also include the baseband processor: MIMO combining scales as $\mathcal{O}(N_r^{2} K)$ MACs per sample, and each MAC at $b$ bits of data width costs $\propto b$ in switching energy (and $\propto b^2$ in multiply energy). For $N_r = 256$ , $K = 8$ , at $f_s = 250$ MS/s, the baseband power is typically several watts — larger than the ADC power for $b \leq 6$ . At very low resolution the baseband can be simplified (MRC collapses to a sign-addition network for $b = 1$ ), recovering an extra factor-of- $b$ savings that analytical figures sometimes ignore. The full-accounting trade curves are surprisingly flat over $b \in \{1,\ldots,4\}$ and the choice of resolution is driven as much by implementation convenience as by the pure rate curve.

🚨Critical Engineering Note

Energy-Rate Trade in mmWave Massive MIMO

mmWave bands (28 GHz, 39 GHz, and the 60-72 GHz sub-THz window) are the case where the low-resolution story matters most. The reasons are structural:

Wide bandwidth ( $\geq 400$ MHz) forces $f_s \geq 800$ MS/s, so the ADC power is $\geq 8\times$ that of a sub-6 GHz link.
High path loss keeps the per-antenna SNR low ( $\leq -10$ dB on the uplink), so Theorem TOptimal Resolution at Low Per-Antenna SNR applies and 1-bit is close to optimal.
Small cell radii mean each BS serves fewer users, so the power saved on ADCs goes straight into battery life or densification.
Hybrid beamforming (Chapter 20) reduces the number of RF chains, so the per-chain ADC cost still dominates the total.

3GPP NR Release 17 contains language explicitly permitting low- resolution ADC architectures for FR2, and commercial equipment typically uses $b \in \{4, 6\}$ to balance rate and power. 1-bit prototypes exist in research testbeds (ETH, Cornell, MIT) but have not yet reached standardization.

Practical Constraints

•
Wideband $f_s$ forces aggressive ADC power reduction
•
Per-antenna SNR below $-5$ dB is typical on mmWave uplinks
•
Baseband processing dominates overall EE above $b = 6$

📋 Ref: 5G-NR FR2, 3GPP TR 38.901; Studer-Durisi testbeds

,

Energy-Efficient ADC Resolution by Operating Regime

Regime	Per-antenna SNR	Optimal $b^\star$	EE vs ideal (relative)
mmWave high-bandwidth uplink	$\leq -10$ dB	$1$	$\times$ 5-10
mmWave typical cell	$-10$ to $-5$ dB	$1$ or $2$	$\times$ 3-5
Sub-6 GHz macro cell	$0$ to $+5$ dB	$3$ or $4$	$\times$ 1.5-2
High-SNR point-to-point	$\geq +10$ dB	$5$ or higher	$\approx$ ideal
XL-MIMO indoor hot-spot	varies across aperture	per-antenna alloc. (Sec. 19.4)	$\times$ 2-3 vs uniform

Why This Matters: ADC Efficiency as a 6G Constraint

The 6G research agenda pushes toward yet higher bandwidths (multi-GHz) and larger antenna counts (thousands), both of which make ADC power the dominant receiver cost. The same energy-rate trade curves of Section 19.5 will be running the design of sub-THz base stations for the next decade, and the interaction with hybrid beamforming (Chapter 20) and RIS (Chapter 21) is where most of the open problems live. The bits-per-joule metric — not bits-per-second — is likely to become the headline spec of the 6G MIMO radio.

Common Mistake: Energy Budget Must Include Calibration and DSP

Mistake:

Energy-efficiency plots that show multi-order-of-magnitude savings from 1-bit receivers typically count only the ADC itself.

Correction:

A complete energy budget has four components: (i) the ADC, (ii) the RF front-end (LNA, mixer, VGA if any), (iii) the baseband MAC array, and (iv) the DC-offset and bias-calibration loops specific to the comparator receiver. At low $b$ , the first vanishes, the second and third contract modestly, and the fourth grows because 1-bit receivers need tighter analog calibration to place the decision threshold at true zero. At $b \geq 4$ the calibration loop is looser (the decision levels are more forgiving). Real savings are 3-10x at best, not the 100-1000x the ADC-only plot suggests.

Key Takeaway

The energy-optimal ADC resolution is a function of operating regime. At low per-antenna SNR (typical mmWave uplink) the 1-bit receiver dominates; at moderate SNR the sweet spot slides to 3-4 bits; at high SNR conventional high-resolution is best. The unifying principle is that $\eta(b) \propto \kappa_b / 2^b$ at ADC-dominated operating points — a sharply unimodal function of $b$ with optimum at $b^\star = 1$ whenever the low-SNR approximation holds, and at $b^\star \in \{3, 4\}$ otherwise.

Mixed-ADC receiver

Massive-MIMO receiver that combines a small fraction $\alpha$ of high-resolution ADCs with a large fraction of 1-bit ADCs. Effective array gain interpolates linearly between $\kappa_1 = 2/\pi$ at $\alpha = 0$ and $1$ at $\alpha = 1$ . Practical deployments use $\alpha \in [0.1, 0.3]$ .

Energy efficiency (bits/J)

Ratio of instantaneous sum rate (bits per second) to instantaneous receiver power consumption (joules per second). The relevant performance metric for ADC resolution choice in massive MIMO, since rate alone would argue for infinite precision while power alone would argue for $b = 0$ .

Quick Check

At very low per-antenna SNR and under the ADC-dominated power model, which statement about the energy-optimal ADC resolution $b^\star$ is correct?

$b^\star$ grows linearly with $N_r$

$b^\star = 1$ (pure 1-bit)

$b^\star = \log_2(N_r)$

$b^\star$ equals the Walden FoM exponent