Fully Digital mmWave MIMO Beamforming

• Fully digital mmWave MIMO beamforming is an architecture where each antenna has its own RF chain and ADC/DAC, enabling fine-grained digital beamforming and spatial multiplexing.
• Key signal processing techniques, including channel-matched SVD, partial zero-forcing, and LMMSE combining, optimize both spectral and energy efficiency.
• Innovations such as low-resolution converters, two-stage combining, and scalable algorithms address hardware constraints while supporting high-capacity mmWave applications.

Fully digital millimeter-wave (mmWave) MIMO beamforming refers to array transceiver architectures in which each antenna element is interfaced with its own independent radio-frequency (RF) chain, analog-to-digital converter (ADC)/digital-to-analog converter (DAC), and digital baseband processing. All beamforming, spatial multiplexing, and spatial domain user separation are performed in the digital domain, offering maximum flexibility and spatial degrees of freedom. The rapid advances in mmWave component integration, low-power data converters, and parallel digital signal processors have elevated fully digital architectures to a central position in modern high-capacity mmWave MU-MIMO systems.

1. System Architecture and Signal Model

A fully digital mmWave MIMO transceiver implements one complete RF chain per antenna, encompassing a DAC and power amplifier on each transmit chain, and an LNA and ADC on each receive chain. Let $N_{\mathrm{T}}$ and $N_{\mathrm{R}}$ denote the number of transmit and receive antennas, respectively, with one RF chain per antenna and a digital baseband block to manage spatial precoding or combining (Buzzi et al., 2016, Buzzi et al., 2017).

The baseband precoder at the transmitter maps $MK$ spatial streams—$M$ streams per user, $K$ users—into $N_{\mathrm{T}}$ digital signals, fed through $N_{\mathrm{T}}$ DACs and subsequently upconverted and amplified. At the receiver, each of the $N_{\mathrm{R}}$ antenna signals is amplified (LNA), downconverted, digitized by an ADC, and digitally combined to recover the $M$ spatial streams per user.

The spatial channel models are typically clustered mmWave MIMO channels, either narrowband or wideband, with a mix of line-of-sight (LOS) and multipath scatterers. The canonical model is

$$\mathbf{H}_k = \gamma \sum_{i=1}^{N_{\mathrm{cl}}}\sum_{l=1}^{N_{{\mathrm{ray}},i}} \alpha_{i,l} \sqrt{L(r_{i,l})} \mathbf{a}_r(\phi^r_{i,l}) \mathbf{a}_t^H(\phi^t_{i,l}) + \mathbf{H}_{\mathrm{LOS}},$$

where $N_{\mathrm{R}}$0, $N_{\mathrm{R}}$1 are receive and transmit array steering vectors, and $N_{\mathrm{R}}$2 is a normalization (Buzzi et al., 2016, Buzzi et al., 2017).

The overall input–output model becomes, with digital precoding/combining,

$N_{\mathrm{R}}$3

with $N_{\mathrm{R}}$4 denoting the digital precoder (collecting all $N_{\mathrm{R}}$5) and $N_{\mathrm{R}}$6 denoting thermal noise.

2. Signal Processing: Beamforming, Precoding, and Combining

Fully digital architectures endow the system with full baseband flexibility: arbitrary (frequency-selective or instantaneous) precoders and combiners can be designed with fine spatial and spectral granularity.

Common digital beamforming/precoding strategies include:

• Channel-matched (CM-FD): Singular value decomposition (SVD) matching, maximizing per-user SNR (Buzzi et al., 2017).
• Partial Zero-Forcing (PZF-FD): Precoding matrices are constructed to null selected inter-user interference by projecting onto the orthogonal subspace of dominant interferers (Buzzi et al., 2016, Buzzi et al., 2017).
• LMMSE Combining: At the receiver, linear minimum mean-square-error (LMMSE) combiners can be designed, even in the presence of nonlinearities and quantization, by adapting the digital combining weights to distortion models (Abdelghany et al., 2019).

The digital implementation allows for fast beam-switching, instantaneous and frequency-selective beam adaption, spatial-division multiple access, and seamless integration with control and initial access signaling (Dutta et al., 2019).

3. Hardware Constraints: Quantization, Nonlinearities, and Energy Model

While fully digital architectures maximize spatial flexibility, they demand a complete RF front end per element, which dominates the power and hardware complexity budget. Key nonidealities include:

Low-Resolution Data Converters: Power and area constraints at mmWave frequencies necessitate the use of ADCs/DACs with reduced resolution (3–4 bits). The additive quantization noise model (AQNM) and the Bussgang decomposition model the quantized signal as the sum of a scaled input and uncorrelated distortion, permitting closed-form analysis of SINR and achievable rates (Abdelghany et al., 2019, Pasic et al., 22 Jan 2026, Dutta et al., 2019).

• With $N_{\mathrm{R}}$7-bit ADCs/DACs, the quantization distortion is characterized by $N_{\mathrm{R}}$8; $N_{\mathrm{R}}$9–4 bits yields an intrinsic SNR $MK$0 dB, with negligible impact on achievable rate for targets up to 10–15 dB SNR per antenna (Abdelghany et al., 2019, Dutta et al., 2019, Pasic et al., 22 Jan 2026).
• Simulation and experimental studies confirm that 4-bit converter resolution replicates the spectral and error bounds of ideal, infinite-resolution digital MIMO links in practical cellular SNR regimes (Pasic et al., 22 Jan 2026, Dutta et al., 2019).

Per-Antenna Nonlinearities: Power amplifier compression and RF chain nonlinearities are modeled with memoryless polynomial functions; system-level models recommend per-antenna input back-off (e.g., 6 dB IBO) to ensure distortion remains below the noise floor (Abdelghany et al., 2019).

Power Consumption Model: The aggregate circuit power includes contributions from every RF chain, ADC, DAC, baseband, LNA, and PA, parameterized as

$MK$1

with typical figures: $MK$2 mW, $MK$3 mW, $MK$4 mW, $MK$5 mW, $MK$6 mW, $MK$7 mW (Buzzi et al., 2016, Buzzi et al., 2017). Power consumption of DACs/ADCs is highly sensitive to quantizer resolution, scaling as $MK$8 ($MK$9 = number of bits) (Dutta et al., 2019, Pasic et al., 22 Jan 2026).

4. Spectral Efficiency and Energy Efficiency

Spectral Efficiency: Fully digital mmWave MIMO enables the realization of the full spatial multiplexing potential, with achievable spectral efficiency (ASE) given by

$M$0

where $M$1 incorporates noise and residual interference (Buzzi et al., 2016, Buzzi et al., 2017). In the large antenna regime, ASE increases as $M$2 per stream, with near-linear user scaling (Buzzi et al., 2017).

Energy Efficiency: The global energy efficiency (GEE), defined as

$M$3

encapsulates the trade-off between data rate and total power expenditure, including both radiated and circuitry power (Buzzi et al., 2016, Buzzi et al., 2017).

Key Findings:

• Fully digital architectures (especially partial zero-forcing) typically deliver the highest spectral and energy efficiency for moderate-to-large arrays and state-of-the-art component power budgets (Buzzi et al., 2016, Buzzi et al., 2017).
• Energy efficiency peaks at moderate array size ($M$4) and/or quantizer resolution ($M$5 bits), with diminishing returns for larger configurations due to rapidly increasing circuit power (Pasic et al., 22 Jan 2026, Buzzi et al., 2017).
• The superiority of FD persists except in extremely large arrays or if phase shifter/RF-chain power is reduced by orders of magnitude, at which point hybrid/analog solutions may become more energy-efficient.

Beamforming	ASE (bit/s/Hz)	GEE (bit/J)
PZF-FD	1680	$M$6
CM-FD	1625	$M$7
PZF-HY	1410	$M$8
AN	1150	$M$9
SW	900	$K$0

Sample at $K$1, $K$2, $K$3, $K$4 dBW (Buzzi et al., 2017).

5. Design Methodologies and Algorithmic Innovations

Low-Resolution Digital Beamforming: Design recipes leverage the AQNM/Bussgang approach to model quantization, enabling efficient optimization of the precoder/combiner under quantization-induced distortion. System-level evaluations recommend 3–4 bit ADCs/DACs for both Tx and Rx, striking a balance between capacity and power (Abdelghany et al., 2019, Dutta et al., 2019, Pasic et al., 22 Jan 2026).

Long-Term Beamforming and Complexity Reduction: For massive arrays ($K$5), explicit channel estimation and instantaneous MMSE combining are computationally infeasible. Scalable methodologies utilize low-rank projections from slowly varying spatial covariance matrices, and polynomial matrix approximations for invert operations, reducing pilot and computation overhead by 10–100$K$6 with negligible SINR loss ($K$7 dB) compared to optimal instantaneous beamforming (Rasteh et al., 12 Nov 2025).

Two-Stage Dimension-Reducing Architectures: In mobile wideband scenarios, two-stage fully digital combining applies a fixed, geometry-adapted combinator to reduce UE baseband streams (updated on the beam coherence time), followed by a second stage combining per-fading realization. This achieves near-ideal spectral efficiency with reduced hardware and processing burden (Khorsandmanesh et al., 6 Aug 2025).

Testbed Prototypes: Real-time fully digital mmWave multi-beam platforms based on RF system-on-chip (RF-SoC) FPGAs demonstrate four simultaneous beams at both 28 GHz (0.8 GHz per beam) and 60 GHz (1.8 GHz per beam), validating practical design trade-offs in ADC/DAC resolution, pipeline timing, and board/radio integration (Pulipati et al., 2020).

6. Practical Performance, Hardware Prototypes, and Implementation Feasibility

Experimental fully digital mmWave platforms have validated core theoretical findings:

• Using four independent RF/ADC/DAC channels at 28 GHz and 60 GHz, per-beam bandwidths of up to 1.8 GHz, with real-time, per-beam digital weighting for multi-beam formation (Pulipati et al., 2020).
• Digital beamforming achieves $K$8 simultaneous multi-beam search, minimizing latency and enabling fast control-plane signaling, in contrast to analog/hybrid solutions which require time-consuming beam-sweeping (Dutta et al., 2019, Pulipati et al., 2020).
• Converter and FPGA resource scaling remain primary limitations for expansion to very large arrays; co-packaged RF front-ends with the SoC and die-level integration are necessary for managing channel counts and SNR at mmWave (Pulipati et al., 2020).

Transmit signal integrity (EVM, ACLR) meets 3GPP NR standards (e.g., $K$9 EVM for 64QAM with 4–6 bit DACs and simple analog filtering), confirming feasibility for both base-station and UE transceivers with low-resolution quantization (Dutta et al., 2019, Pasic et al., 22 Jan 2026).

7. Trade-Offs, Limitations, and Future Directions

Trade-Offs:

• FD architectures offer unmatched flexibility and maximal spatial DoF but incur a power/complexity penalty for large arrays.
• Low-resolution quantization and two-stage combining architectures offer tractable complexity/power scaling while maintaining most of the FD performance advantages (Pasic et al., 22 Jan 2026, Khorsandmanesh et al., 6 Aug 2025).
• For network operators, the array size and quantizer resolution must be co-optimized for both spectral and energy efficiency; 4-bit converters and moderate array sizes are widely corroborated as the energy-spectral sweet spot (Buzzi et al., 2017, Pasic et al., 22 Jan 2026).

Limitations:

• FD designs remain sensitive to power consumption in very large arrays or where advanced analog processing can be implemented with ultra-low power per element.
• Channel state information (CSI) acquisition and synchronization challenges remain, particularly in FDD and mobile non-reciprocal environments.

Outlook:

• Future improvement in RF front-end, baseband integration, and mmWave digital converter FoM will further extend the regime of FD superiority in energy efficiency and capacity.
• Algorithmic advances in long-term beamforming, covariance-aware designs, and hardware-aware signal processing will be critical for realizing the full performance potential at ultra-massive scales (Rasteh et al., 12 Nov 2025).
• Hybrid quantization, adaptive-resolution arrays, and reconfigurable RF architectures present avenues for further optimization.

Fully digital mmWave MIMO beamforming has transitioned from an aspirational technology to a deployable, competitive solution for high-capacity wireless systems, supported by both theoretical analysis and hardware validation across the mmWave spectrum (Buzzi et al., 2016, Buzzi et al., 2017, Dutta et al., 2019, Abdelghany et al., 2019, Pulipati et al., 2020, Rasteh et al., 12 Nov 2025, Khorsandmanesh et al., 6 Aug 2025, Pasic et al., 22 Jan 2026).