TL;DR
As 6G moves into the "upper mid-band" (7–24 GHz), base stations are becoming massive hardware undertakings with thousands of antennas. This paper introduces the first detailed parametric power model for these FR3 base stations, revealing that at high antenna counts, the power consumed by digital and analog signal processing actually eclipses the power used by the amplifiers themselves.
What Problem Does This Paper Address?
The transition to 6G requires massive jumps in data capacity, leading researchers to the "Frequency Range 3" (FR3) band. This band offers a "Goldilocks" zone: more bandwidth than traditional sub-6 GHz bands but better coverage than millimeter-wave (mmWave).
However, realizing this potential requires Base Stations (BS) with massive MIMO arrays—sometimes thousands of antenna elements. The industry faces a major sustainability hurdle: how much power will these monsters actually draw? Existing power models for older generations don't account for the specific hardware architectures (like hybrid beamforming) or the high-speed processing required for FR3. The authors aim to fill this gap with a model that helps engineers trade off performance against energy costs.
Methods
The authors developed a parametric power model that breaks a base station down into three primary "buckets," further subdivided by their operational state (data transmission, signaling, micro-sleep, or idle):
- Digital Signal Processing: Modeled using "Giga Operations Per Second" (GOPS). They distinguish between components better suited for ASICs (like encoders) and those requiring the flexibility of FPGAs (like precoders).
- Analog Signal Processing: This includes the power draw of Digital-to-Analog Converters (DACs), mixers, local oscillators, and phase shifters. They use "Figure of Merit" (FoM) models to estimate power based on sampling rates and resolution.
- Power Amplifiers (PAs): Instead of assuming constant efficiency, they use a non-linear model that accounts for "back-off" (running the PA below its peak to maintain signal quality).
The model specifically compares Fully-Digital Beamforming (one RF chain per antenna) against Hybrid Partially-Connected Beamforming (sharing RF chains across subarrays of antennas).
Datasets
This is a modeling and simulation study rather than a training-based machine learning paper. To evaluate their model, they used:
- Sionna Simulator: Used to generate a 3GPP 39.801 Urban Micro (UMi) wideband channel model.
- Hardware Benchmarks: Power and efficiency data were derived from the AMD Zynq UltraScale+ RFSoC ZCU216 and various industry datasheets for mixers, LNAs, and filters.
Key Insights and Ideas
- The "Processing Pivot": In traditional base stations, the Power Amplifier is the primary power hog. However, in FR3 systems with many RF chains (512+), the digital and analog processing components take over, consuming 2× to 4× more power than the PAs.
- The Case for Hybrid: Fully digital beamforming is often seen as the "gold standard" for performance, but the paper shows it is incredibly power-hungry. Hybrid beamforming provides a much better balance, maintaining high data rates while significantly cutting the hardware overhead.
- Load Sensitivity: The model accounts for "micro-sleep" modes. It highlights that at realistic traffic loads (around 30%), the static power consumption of the hardware is a dominant factor, making hardware efficiency just as important as clever software algorithms.
Key Results
- Energy Efficiency (EE): Hybrid beamforming with 256 RF chains (for a 1024-antenna array) achieved the highest energy efficiency, reaching 13.6 Mbit/J at full load.
- Efficiency Gains: Compared to a fully-digital architecture, the hybrid approach improved energy efficiency by 1.78× at a 30% resource load.
- Throughput: The proposed FR3 setup achieved a downlink sum rate of over 13 Gbit/s (roughly 1.3 Gbit/s per user) using a 400 MHz bandwidth.
- Power Scaling: For a 1024-antenna array at 30% load, total power consumption ranged from 300W (with 16 RF chains) to over 1000W (fully digital).
Limitations and Future Work
- TDD Assumption: The model assumes Time-Division Duplexing (TDD) and channel reciprocity, which simplifies the signaling overhead. Frequency-Division Duplexing (FDD) would likely introduce more complexity.
- Specific Scenario: The results are tied to an Urban Micro (UMi) environment; performance might vary significantly in indoor or high-mobility outdoor scenarios.
- Hardware Evolution: As ASIC and FPGA technology improves (e.g., smaller nanometer processes), the "Processing Pivot" point will shift, requiring regular updates to the model's parameters.
Summary updated