Network energy efficiency is usually presented as a cost topic. I read it that way for a long time, and it explained nothing. The useful question is not what the electricity costs. It is where the electricity goes, and which part of it can be removed at all. Electricity is a large operating cost for a mobile operator, and the recent energy crisis made that cost visible. The structure below is the one the NGMN Green Future Networks programme settled on, in its Network Energy Efficiency Phase 2 work. The available options sort into three groups, and each group has its own time horizon. The first group is process optimisation, and it is short term. The second is engineering optimisation, and it is medium term. The third is new technologies, and it is long term. The three groups differ in one important way. Process optimisation configures hardware that is already installed. Engineering optimisation replaces that hardware. New technologies are not yet deployed at scale. Therefore the saving per group rises, and so does the lead time and the capital cost.

If you take one number away from this page, take this one. The RAN consumes roughly three quarters of the electricity used by a mobile network. So a percentage saved in the RAN is worth about three times a percentage saved anywhere else. That is why almost every solution in this note points at the radio site.

• Executive Summary
• Acronym
• Why is network energy efficiency an engineering problem, and not only a cost problem?
• Where does the energy actually go in a mobile network?
• Why does the static and dynamic split decide what can be saved?
• How are the options grouped, and why by time horizon?
• Which network energy saving features does 3GPP already specify?
• Why is turning the features on not enough on its own?
• How is the energy efficiency of a site actually measured?
• How much does data-driven optimisation save in a live network?
• How can the radio unit itself be made more efficient?
• What can be done about the baseband unit?
• How much energy is lost in the passive antenna before the signal leaves the site?
• Why are DC power distribution losses growing, and what fixes them?
• How much energy does cooling consume, and what does liquid cooling change?
• Does network disaggregation save energy, or spend it?
• What can advanced radio technologies add?
  • What does a reconfigurable intelligent surface offer?
  • What does distributed massive MIMO offer?
• What does the prioritised list of solutions look like?
• How should these percentages be combined?
• Where did this work go next?
• Reference

Executive Summary

Acronym

This note uses the vocabulary of two different worlds. One is the radio world, and the other is the power and cooling world. The table below covers both. Where a term carries a specific meaning in this report, that meaning is given as well.

Why is network energy efficiency an engineering problem, and not only a cost problem?

Electricity is a finite resource, and it is a significant operating input cost for a mobile network. That much is a commercial statement, and it is where most presentations stop. The engineering statement is the one I find useful. Most of the electricity a network consumes is not spent carrying traffic. It is spent keeping equipment ready to carry traffic. That distinction is what makes the problem tractable at all.

Consider what happens at a lightly loaded cell. The power amplifiers stay biased. The transceiver chains stay powered. The cooling system keeps running. The DC feed keeps losing energy in the cable. Meanwhile very little data is being sent. Therefore the energy per delivered bit at low load is far worse than at high load. So the largest single opportunity is not to make transmission cheaper. It is to stop paying for readiness that nobody is using.

This is why the report groups the solutions the way it does. Process optimisation reduces readiness cost by switching things off more aggressively, using features that already exist. Engineering optimisation reduces readiness cost by installing hardware whose idle draw is lower. New technologies change the shape of the problem, for example by removing the cooling load or by improving the channel itself. Each group attacks a different part of the same waste.

• The waste is idle power, not transmit power : Transmit power is a small part of a base station's consumption. The larger part is the fixed cost of being switched on. That is why sleep modes, shutdown features and idle-draw reductions dominate the solution list.
• The RAN is where the energy is : The RAN accounts for almost three quarters of the electricity consumed by a mobile network. Therefore a solution that saves 10% in the RAN outranks a solution that saves 30% in the core.
• Three groups, three time horizons : Process optimisation is short term. Engineering optimisation is medium term. New technologies are long term. The grouping is not about how clever the idea is. It is about how long it takes to deploy and how much capital it needs.
• The energy crisis changed the urgency, not the physics : The options were available before. What changed is the price of not using them.

Where does the energy actually go in a mobile network?

Start with the split, because the split decides the ranking. Everything else on this page sits downstream of the three numbers below. The first level is the network itself. Almost three quarters of all electricity consumed by a mobile network is consumed at the mobile sites. The remaining quarter sits in the core, in transport and in the data centres.

The second level is inside the RAN. Base station equipment accounts for at least 50% of RAN energy. Cooling accounts for up to 40%, and that share depends heavily on the site. An indoor site in a hot climate is not the same as an outdoor cabinet in a cold one. Everything else accounts for about 10%.

The third level is inside the base station equipment. The RU accounts for about 80%, the BBU for about 13%, and the main control for about 7%. A separate measurement across a whole national network sharpens that further. There the RU reaches 88% of base station equipment energy in 5G, against 82% in 4G. Note that this second study does not include the main control, so its two numbers are shares of RU plus BBU only.

Multiplying the levels gives the number that matters. The RU is about 80% of base station equipment, and base station equipment is about 50% of RAN energy. Therefore the RU alone is about 40% of RAN energy. That single figure explains the priority order of everything that follows.

Figure 1. The shares multiply. The Radio Unit ends up as roughly 40% of all RAN energy, which is why every ranked list in this note starts there.

• Three quarters of the electricity is in the RAN : This is the first level, and it decides where an energy programme should look. The core and the transport network matter, but they are the smaller quarter.
• Cooling is the surprise item : Up to 40% of RAN energy is spent removing heat, not producing radio. That share varies strongly with site type, climate and whether the site is indoor or outdoor.
• The Radio Unit is about 40% of RAN energy : This follows from 80% of the base station equipment, which is itself at least 50% of the RAN. So a saving in the RU is worth roughly twice a saving anywhere else in the base station.
• The RU share is growing with each generation : It measures 88% of RU plus BBU energy in 5G, against 82% in 4G. The reason is that AAUs added RF chains, bandwidth and transmit power, while baseband processing scaled more slowly.
• Check what a share is a share of : The 80% figure includes the main control, and the 88% figure does not. Reading them as the same quantity is the most common mistake with these breakdowns.

Why does the static and dynamic split decide what can be saved?

Knowing that the RU is 40% of RAN energy is only half the picture. The other half is what that energy is spent on, and this is the part I had wrong. I assumed most of it was transmit power. It is not. RU energy consumption splits into two parts, and they behave completely differently.

The first part is static energy consumption. It does not depend on the traffic that the RU carries. Its main driver is the number of RF transceiver chains, because that count sets the number of power amplifiers. The second part is dynamic energy consumption. It does depend on traffic load, and it also depends on the maximum transmit power of the equipment.

The split differs sharply between equipment types. In an RRU, static consumption is about 35% of the total. In an AAU, it is about 56%. The reason is the hardware that massive MIMO needs. Beamforming and spatial multiplexing require many more transceiver chains, and each of those chains draws power whether or not it is carrying data.

The absolute numbers make the same point more bluntly. A 5G RU with 100 MHz of bandwidth, 64 antennas and 240 W maximum transmit power consumes about 4077 W. A 4G RU with 20 MHz, 8 antennas and 40 W maximum transmit power consumes about 950 W. The 5G unit has roughly eight times the RF channels, five times the bandwidth and a much larger transmit power. In addition, 5G AAU consumption varies widely even between units with the same antenna count. That variation is not seen in 4G RRUs.

So the two parts need two different tools. Static consumption is attacked by switching hardware off, which is what the shutdown and sleep features do. Dynamic consumption is attacked by carrying the same traffic with fewer resources, or at lower transmit power. A high static share is bad news for the energy bill at low load. However, it is good news for the shutdown features, because there is more to switch off.

Figure 2. An AAU spends more of its energy on simply being switched on than an RRU does. Therefore the shutdown features have more to work with, and the low-load penalty is worse.

• Static consumption is driven by RF chain count : Each transceiver chain needs a power amplifier, and that amplifier draws power regardless of traffic. So the antenna count on the datasheet is also an idle-power figure.
• An AAU is 56% static, an RRU is 35% : The extra hardware for massive MIMO beamforming and spatial multiplexing is what moves the split. That is the price of the capacity and coverage gain.
• A high static share cuts both ways : It makes low-load operation expensive, which is the problem. However, it also makes shutdown features effective, which is the remedy.
• The absolute gap between 4G and 5G is large : About 950 W against about 4077 W in the quoted comparison. Note that these are not like-for-like units, because the 5G unit has five times the bandwidth and eight times the RF channels.
• Equipment of the same class is not equally efficient : 5G AAUs with identical antenna counts show a wide spread in power consumption. Therefore procurement, and not only configuration, is an energy lever.

How are the options grouped, and why by time horizon?

Every solution on the list falls into one of three groups. The grouping looks like a taxonomy of technique. In practice, it is a taxonomy of deployment effort. That is the more useful reading.

The first group is process optimisation, and it is short term. Nothing physical changes. What changes is the configuration of features that the equipment already supports. Therefore the only cost is software, engineering time and the risk of degrading service quality. The second group is engineering optimisation, and it is medium term. Here the hardware changes. Radios, antennas, cables and power distribution are replaced or re-designed. The cost is capital, and each change needs a site visit. The third group is new technologies, and it is long term. These are not yet deployed at scale. Several of them still need standardisation or an ecosystem before an operator can buy them.

The gains rise across the three groups, but so does the lead time. That is the trade the grouping is really describing.

Figure 3. The grouping is really about deployment effort. The saving per item rises from left to right, and so does the time and the capital needed to obtain it.

• Process optimisation changes nothing physical : It only changes how existing features are configured. Therefore it is the natural starting point, and it is the only tier that can move within a quarter.
• Engineering optimisation needs a site visit : Radios, antennas, jumpers and DC cabling are replaced. So these items should be attached to a planned refresh or to a 5G build, rather than run as a separate campaign.
• New technologies are not yet purchasable at scale : Several of them still need standardisation, an ecosystem or equipment redesign. They belong in a medium to long term plan, and not in this year's target.
• The percentages have different denominators : 3% is a share of total network energy. 30% per RRU is a share of one unit's energy. 60% is a share of DC distribution loss only. Reading them as comparable is the easiest error to make with this table.

Which network energy saving features does 3GPP already specify?

The short-term tier depends on features that already exist in the specification. So they are worth naming one by one. All of them are available today, and none is future work.

Earlier releases concentrated on UE power saving. From 5G-Advanced onwards the focus moved to the network side as well. The RAN was identified as the primary contributor to energy use. Therefore the aim of the network energy saving work is to make more efficient use of radio resources at low and medium load.

3GPP defined a base station power consumption model for evaluation purposes, and captured it in TR 38.864. That model matters more than it first appears. It describes the power consumed per slot for different configurations, and it separates the sleep states from the active state. Without an agreed model, two companies could report two different savings for the same technique and both be right. With the model, techniques can be ranked against each other.

Four families of technique come out of that work. The first is time domain adaptation. Cell DTX and cell DRX let a cell stop transmitting and stop receiving for defined periods, so the hardware can enter a sleep state. The second is frequency domain adaptation, which includes energy saving bandwidth operation and SSB-less SCell operation. The point of SSB-less SCell operation is that an SCell no longer has to transmit its own always-on synchronization burst. The third is spatial domain adaptation, which dynamically changes the number of active massive MIMO antenna ports and transceiver chains. That attacks the static consumption identified in the previous section. The fourth is power domain adaptation, which adjusts PDSCH transmit power to what is actually needed.

Alongside these, AI/ML for NG-RAN is being studied so that cell load and cell state can be predicted. A prediction enables actions such as traffic offloading and dynamic sleep, taken before the load arrives rather than after. Work on energy efficiency as a service criterion is also running at the system level, beyond the RAN.

The common thread is that all of these techniques switch hardware off. The design constraint is therefore the same in every case. The saving must not damage UE performance or network performance while the hardware is off.

• The power model is the enabling piece : TR 38.864 defines base station power per slot, and it separates sleep from active. Without it, saving claims from different sources cannot be compared.
• Cell DTX/DRX is the time domain tool : The cell stops transmitting and receiving for defined periods, so hardware can sleep. The depth of that sleep is limited by whatever must stay always-on, and SSB is the usual limit.
• Spatial adaptation attacks the static part directly : Reducing the number of active antenna ports and transceiver chains reduces exactly the consumption that does not depend on traffic. That is why it is effective on AAUs in particular.
• SSB-less SCell operation removes an always-on cost : An SCell that does not transmit its own synchronization burst can sleep for longer. This is a frequency domain measure with a time domain benefit.
• AI/ML is being studied to decide when to act : Predicting cell load and cell state allows sleep and offloading decisions to be taken in advance. The feature set is the mechanism, and the prediction is the policy.
• Every one of them trades performance for energy : That is the whole difficulty. The features are not free, and the next section is about what that costs in practice.

Why is turning the features on not enough on its own?

If I had to reduce this whole page to one sentence, it would be the next one. Simply switching the features on in the network is not enough. Two things have to be true first. The impact of each feature on energy consumption must be understood. In addition, each feature must be correctly configured, so that end-user performance is not compromised.

Understanding the impact means modelling the RU. The 3GPP model gives a per-slot description of power consumption for different configurations. A second class of model works the other way round. It is fitted to measurements, so it captures the relationships between the components that actually drive RU power in a live network.

This is where machine learning enters. An ML framework can be trained on data gathered from many RU types, with different hardware configurations, and potentially from different vendors. The fitted analytical model then describes RU power consumption during running time, including the effect of the power saving features. One property of this approach is worth stating clearly. Because the ML model generalises, it can cover configurations that were never directly observed in the collected data. That is exactly what is needed, because no operator can measure every combination of feature settings on every RU type.

Configuring correctly is the harder half. A shutdown threshold that is too aggressive saves energy and loses throughput. A threshold that is too conservative protects throughput and saves nothing. Furthermore, the features interact. One example from the report makes this concrete. Where 4G and 5G share a wideband power amplifier, deactivating one cell may save no energy at all. The reason is that the wideband PA has to stay active to serve the other co-deployed cells on the same RF chain. So a per-cell view of energy saving can be simply wrong.

• A feature that is on is not a feature that is tuned : The saving lives in the thresholds, the time windows and the cell pairings. Enabling the feature is the beginning of the work, not the end of it.
• Model the RU before configuring it : A per-slot power model plus a data-fitted model together predict what a setting will actually save. Otherwise the configuration is guesswork with a plausible name.
• ML generalises to configurations never measured : This is the practical reason to use it. The number of feature and hardware combinations is far larger than any measurement campaign can cover.
• Shared wideband power amplifiers break per-cell reasoning : Where one wideband PA per RF chain serves several co-deployed cells, shutting one cell down may save nothing. The PA has to stay on for the others.
• Performance is the constraint, not an afterthought : Every saving in the next section is quoted together with the KPIs that stayed stable. A saving reported without its KPIs is not a result.

How is the energy efficiency of a site actually measured?

Before a network can be optimised, its sites have to be comparable. Energy per bit is the obvious metric, and it is also misleading. A site that delivers a lot of bits slowly can look efficient, while a site that delivers fewer bits quickly can look wasteful. However, the second site may be the one meeting its service target.

The way around this is a QoS-adjusted metric. Energy efficiency is defined as the data volume divided by the energy consumption, and then multiplied by a rate factor. The rate factor is the average downlink rate divided by the average target downlink rate. In this way, a site is rewarded for delivering volume cheaply and penalised for missing its rate target. Note that the rate factor is one choice among several. A QoS factor could equally be built on coverage, on delay, or on a combination.

The metric was computed for 168 base station sites in a European metropolitan area. Each site was then plotted against a benchmark curve, so that low-efficiency sites could be identified rather than guessed at. The low-efficiency sites fell into three types, and each type needs a different remedy.

Type I sites have low traffic and a high downlink rate. There the energy saving features are not fully used, so the remedy is to activate them more aggressively. Type II sites have high traffic and a low downlink rate. There the features are already fully used, so the remedy is a hardware refresh. Type III sites have high traffic and a high downlink rate, and their hardware is new. There the remedy is to migrate traffic from LTE onto NR. Results are available for the first two types. The third was left for later work.

• Energy per bit alone ranks sites wrongly : It rewards slow delivery. Therefore the metric has to carry a quality term, or the optimisation will quietly trade away user experience.
• The QoS factor is the rate against its target : Data volume divided by energy, multiplied by achieved rate over target rate. Coverage or delay could be used in the same position instead.
• A benchmark curve turns a metric into an action list : Plotting 168 sites against current and future benchmark curves identifies which sites are underperforming, and by how much.
• Three site types, three different remedies : Low traffic with a high rate calls for more aggressive feature activation. High traffic with a low rate calls for a hardware refresh. High traffic with a high rate and new hardware calls for traffic migration to NR.
• Diagnosis comes before treatment : The same energy saving feature is the right answer at a Type I site and the wrong answer at a Type II site. So the classification step is what makes the campaign efficient.

How much does data-driven optimisation save in a live network?

Two trials give numbers, and they measure different things. You will see the figures quoted as though they were one result. I mixed them up on a first reading, so this section keeps them apart.

The first trial acted on the Type I sites identified above. Optimised activation of the energy saving solutions saved 1473.5 kWh per day against the baseline. The KPIs for network performance and UE service quality stayed stable. Overall, the QoS-adjusted energy efficiency metric for those sites improved by 17.5%. Note that 17.5% is an improvement in the metric, and not a reduction in energy.

The second trial acted on Type II sites, and it replaced hardware. Two sites had their RUs and BBUs swapped for more efficient multi-band equipment. Energy per site fell from 63 kWh to 37 kWh, which is a 41% saving. At the same time, the UE rate rose from 14 MBps to 28 MBps, which is a doubling. Because the metric multiplies a lower energy by a higher rate factor, the QoS-adjusted energy efficiency improved by up to 260%. That figure is real, but it is the product of two effects and not a 260% energy saving.

The third and most quoted result comes from a separate 4G and 5G network in a large metropolitan area. Here the 4G and 5G carriers were served from the same radio unit, through wideband power amplifiers. That makes the optimisation harder, for the reason given in the previous section. The trial ran in three phases.

In Phase I all power saving solutions were deactivated, so that a true baseline could be measured. That includes symbol, RF and carrier shutdown, and deep dormancy. In Phase II the shutdown solutions were tuned and activated using a traditional rule-based approach. That approach relied on system-level simulation and on frontline engineering experience, and the features ran in a fixed 0 to 6 hour window. In Phase III the data-driven methodology set the same parameters. It optimised the coverage and capacity cell pairing, the shutdown thresholds, and the periods of the day in which each solution may operate. It also optimised the related handover parameters.

The results follow directly. The rule-based approach saved 245 kWh per day, which is a 3% saving against the baseline. The data-driven approach saved 1000 kWh per day, which is a 12% saving against the same baseline. Against the rule-based approach, the data-driven approach saved a further 755 kWh per day, which is 9.1%. Meanwhile the average UE rate and the share of users in good coverage, above -105 dBm, were only minimally affected.

Two of those numbers turn up again in the recommendations table, and the difference is worth resolving. That table lists rule-based automation at 3% and AI-based automation at 9%. So the 9% figure is the incremental gain of the data-driven approach over an already rule-based network. The 12% figure is the total gain against a network with the features switched off. Both are correct, and they simply have different baselines.

Figure 4. Rule-based activation captures a quarter of what the same features can deliver. The rest comes from tuning the thresholds, the operating hours and the coverage-capacity cell pairing.

• Rule-based activation captured 3% : That is 245 kWh per day against a baseline with all the features disabled. It is a real saving, and it is a quarter of what the same hardware can give.
• Data-driven optimisation captured 12% : That is 1000 kWh per day against the same baseline. The extra came from tuned thresholds, optimised operating hours, cell pairing and handover parameters.
• 9% and 12% are the same result with different baselines : The 9.1% is the step from rule-based to data-driven. The 12% is the step from all-features-off to data-driven. Always state which baseline a saving is measured against.
• A hardware refresh gave 41% at the site level : Energy per site fell from 63 kWh to 37 kWh, and the UE rate doubled from 14 to 28 MBps. The reported 260% is the improvement in the QoS-adjusted metric, and it combines both effects.
• 17.5% is a metric improvement, not an energy saving : The Type I trial saved 1473.5 kWh per day and improved the QoS-adjusted metric by 17.5%. These are two different quantities in the same sentence.
• The KPIs were reported alongside every figure : Average UE rate and the share of users in good coverage stayed stable. Without that, a 12% saving would not be a usable result.

How can the radio unit itself be made more efficient?

The RU is about 40% of RAN energy, so this is where the engineering tier starts. Three levers are available, and they work in different ways.

The first lever is band consolidation, and it is the one most often misread. A site that has to serve three bands can do it two ways. It can use three single-band RRUs, or it can use one tri-band RRU. The second arrangement is reported at around 30% less energy for the same three bands. Note carefully what is being compared. A tri-band unit draws more than a single-band unit, so this is not a saving per box. It comes from removing the duplication between boxes. Three single-band units each carry their own static consumption, and one tri-band unit carries it once. Three design choices make that possible. The first is a larger number of supported transmit and receive chains. The second is multi-band power amplifiers. The third is letting those amplifiers share transmit power across the bands they support.

The second lever is the transmit chain count, and it works indirectly. Going from 2T to 4T adds diversity gain. Going to 8T enables beamforming and a sector split. Neither of those saves energy by itself, because more chains means more static consumption. However, the gains allow the same coverage and capacity at a lower transmit power. The total effect reaches as much as a 60% reduction in transmit power. That reduces RU energy consumption significantly.

The third lever applies to AAUs, and it looks wrong at first sight. Adding antenna elements sounds like adding power. It is not, because antenna elements are passive. In an AAU the number of power amplifiers follows the number of transceiver chains. It does not follow the number of radiating elements. Several elements are driven by one amplifier. So doubling the elements per amplifier leaves the amplifier count untouched. What changes is the array. Twice the elements in the vertical direction is twice the vertical aperture, and that is worth about 3 dB of extra antenna gain. Those extra elements draw no current of their own. Therefore the amplifier can be backed off by the same 3 dB, and the coverage stays where it was. Trials with double the number of elements in the vertical direction report up to 30% energy saving per AAU under realistic network and traffic conditions. The cost is a narrower vertical beam, so the elevation coverage of the sector has to be checked.

Notice that the second and third levers aim at the same target from opposite ends. Both buy link budget, so that transmit power can come down. The difference is what they spend to get it. More transmit chains add amplifiers, and therefore add static consumption. More elements per amplifier add none. That is why the third lever is the cheaper of the two in energy terms. Amplifier power is expensive, and passive gain is nearly free.

• Tri-band RRUs remove duplicated static consumption : One unit with multi-band power amplifiers replaces three single-band units. The reported saving is around 30% for the three bands together. It is not a saving per box, because the tri-band unit draws more than any single unit it replaces.
• More transmit chains cut transmit power, not chain power : 4T adds diversity and 8T adds beamforming and sector split. The reported effect is up to a 60% reduction in transmit power for the same service.
• Antenna elements are cheaper than amplifier watts : The amplifier count follows the transceiver chains, and not the elements. So doubling the elements per amplifier adds no amplifiers. Twice the vertical aperture is about 3 dB of extra gain, and the amplifier is backed off by the same amount. The reported saving is up to 30% per AAU.
• That gain is not entirely free : A taller array has a narrower vertical beam. So the elevation coverage of the sector has to be checked. The physical size and the wind load of the unit both grow as well.
• Sharing transmit power across bands helps : A multi-band PA can move power between its bands. Therefore it is sized for the combined peak, and not for the sum of the individual peaks.
• These are refresh-cycle items : Every one of them requires new hardware on the tower. Therefore they are planned against equipment end-of-life, and not against an annual energy target.

What can be done about the baseband unit?

After the RU, the BBU is the next stop, and it is a much smaller one. It sits at around 13% of base station equipment energy. Its consumption splits into static and dynamic parts, in the same way as the RU. That immediately suggests one structural remedy. If a single BBU serves several cells, the static consumption is pooled across those cells, so the cost per cell falls.

The measured picture is less tidy than that. BBU energy consumption per cell varies widely between BBU types, even under normalised conditions. Two reasons are given. The first is the number and the type of features implemented, because each feature adds digital signal processing and therefore energy. The second is the hardware base. A BBU built on a general-purpose server has a different power consumption profile from one built on dedicated components.

Two directions for improvement are named. The first is deep dormancy, where available, so that the BBU can be switched off when a cell has no traffic. The second is network function virtualisation. Disaggregating the BBU into a CU and a DU allows the CU resources to be pooled in a data centre and shared between several cell sites. Software features such as cell off, channel card off and model off also reduce consumption directly.

Nobody has published data on how much these approaches actually save, and the topic is flagged for further study. That gap is worth carrying forward, because a virtualised BBU is often assumed to be more efficient rather than measured to be.

• The BBU is about 13% of base station equipment energy : It is worth attention, but a 30% saving there is worth under 2% of RAN energy. So it ranks below the RU by roughly a factor of six.
• Pooling spreads the static cost : One BBU serving several cells amortises its traffic-independent consumption. That is the clearest structural saving available.
• BBU types differ widely under identical conditions : Feature count and hardware base both drive the spread. Therefore, as with radios, procurement is an energy decision.
• Deep dormancy and CU/DU pooling are the named remedies : Dormancy removes the idle draw when a cell is empty. The CU/DU split moves the poolable part into a shared data centre.
• No saving figure is published : The extent of these reductions is unknown, and it is explicitly left for further study. An assumed saving is not a measured one.

How much energy is lost in the passive antenna before the signal leaves the site?

A passive antenna consumes no electricity at all. I nearly skipped this section for that reason. That would have been a mistake, and the reason is simple. Whatever the antenna loses as heat has to be produced again by the power amplifier. So an antenna loss is an amplifier cost.

Three loss types are named. Return losses come from the impedance mismatch between the RU and the antenna. Insertion losses are the energy lost as heat inside the antenna. Coverage losses come from the difficulty of directing RF energy at the region that needs it, for example because of side lobes. The overall measure that combines these is antenna radiation efficiency, defined as the energy radiated by the antenna divided by the energy delivered to it.

The largest single item is the feeding network. In a traditional design, the RF signal travels from the antenna port through phase shifters, power splitters and combiners, all connected by coaxial cables, before it reaches the dipole array. That path dissipates energy, and the amount depends on the band. It is typically about 25% below 1 GHz, and about 35% between 1.7 and 2.6 GHz. A typical configuration with 2 low bands and 4 high bands therefore loses about 30%.

Simplifying the feeding network cuts that figure to about 17%. Three changes do the work. The distribution components are combined into one. The number of component connections is reduced. Coaxial cables are replaced with air-type strip lines. The result is close to a 50% reduction in feeding path loss.

The remaining items are smaller, but they are not negligible. Die-cast dipoles lose under 1%, a high-quality substrate loses about 1.5%, and a poorer substrate such as FR4 epoxy can lose up to 10%. A fiberglass radome loses about 5%, and new materials can halve that to about 2%. RF cabling adds its own share. At 2.1 GHz the loss over a 4-metre jumper cable is 11%. Redesigning the site so the RRU sits close to the antenna can shorten that jumper to around 0.3 metres. On 5G sites carrying both FDD and TDD bands, cluster connectors give the lowest RF path insertion loss.

One further option addresses coverage loss rather than insertion loss. On sites carrying both 4G and 5G, the RRU and the AAU can be integrated into a single unit mounted behind the passive antenna. Placing that hybrid unit at the same height as the passive antenna improves the coverage of the radio being added. Better coverage means less energy is needed to serve the same area.

Figure 5. The feeding network is the dominant loss in a passive antenna, and it is also the one with the largest available improvement.

• A passive antenna draws no power but still costs energy : Everything it loses as heat has to be generated again by the power amplifier. So antenna efficiency is an amplifier saving in disguise.
• The feeding network is the dominant loss : About 30% for a typical 2 low band and 4 high band configuration. Simplifying it brings that to about 17%, which is close to a 50% reduction in the feeding path loss itself.
• Substrate choice can cost 10% : Die-cast dipoles lose under 1%, a high-quality substrate about 1.5%, and FR4 epoxy up to 10%. This is a specification decision, not an operational one.
• Jumper length is a real number : 11% is lost over a 4-metre jumper at 2.1 GHz. Mounting the RRU close to the antenna shortens that to around 0.3 metres.
• Coverage loss is fixed by placement, not by materials : Integrating the RRU and AAU into one unit at the same height as the passive antenna improves coverage. Less coverage loss means less transmit power for the same area.

Why are DC power distribution losses growing, and what fixes them?

This is the least visible item on the whole list, and possibly the most striking. The energy discussed here never reaches the radio at all. It is lost in the cable between the power source at the base of the tower and the radios at the top.

The cause is historical. Up to about 2008, 1G and 2G networks kept the RF amplifiers in the cabinet at the base of the tower. Large coaxial cables then carried the RF signal up to the antenna. With 3G the RRU concept moved the amplifiers up the tower, so DC power had to be fed upward instead. Based on the RRU power requirements of the time, most operators used a #8 AWG multi-conductor trunk cable per RRU. That practice became the de facto standard, and it is still the default today.

The problem is that the assumption behind it no longer holds. More RRUs and then AAUs were added, for new bands and for higher capacity. So more #8 AWG trunk cables were run up the tower, one at a time and on demand. Meanwhile RU power consumption grew, driven by wider channel bandwidths, higher transmit power, carrier aggregation, higher-order MIMO and active beamforming.

Now the physics takes over. DC power loss follows the I²R law, so it grows with the square of the current. For a given conductor size, doubling the DC current quadruples the DC power loss. Therefore loss grows faster than RU power does. If the industry keeps building sites with the current design, the report expects DC power losses to rise by a factor of 4 to 10.

The proposed remedy borrows from the DC power bus systems used in central offices. Instead of running an individual feed for each RU from the base to the top, a bulk feed DC bus carries the power as close to the RUs as possible. A DC distribution box at the tower top then supplies each RU from that bus. The box should sit as close to the RUs as possible, so that the efficiency of the bus is not given back in the last few metres. Two practical points follow. The distribution box can be controlled remotely over TCP/IP or Modbus RTU, which helps operations and enables far more effective load shedding during an outage. In addition, aluminium can be used for the bus feeds. For the same ampacity, aluminium is typically one third of the cost and one half of the weight of copper.

A worked example makes the size of the gain concrete. A single sector carries 8 RRUs, at a distance of 60 m from the DC source. The float voltage is 54 V, and the peak RRU power is 1000 W. Take the benchmark first, that is one #8 AWG trunk cable per RRU. The current is 18.52 A, and the voltage drop is 5.36 V. So the loss is 99.18 W per RRU at peak. Now take a dual feed A/B bulk bus of 250 MCM aluminium. The current per branch is 74.07 A, and the voltage drop is 2.13 V. So the loss falls to 39.51 W per RRU. That is a reduction of 59.67 W per RRU, or 60%.

At an electricity rate of 0.13 $/kWh, 59.67 W saved continuously is 523 kWh and about $68 per RRU per year. The per-site figure comes to about $1,631 per year. That corresponds to 24 RRUs, or three sectors of eight. At 50% average RRU power the same calculation gives 131 kWh and about $17 per RRU per year. The percentage reduction stays at 60% in both cases, because both designs scale with the same square law.

Figure 6. Splitting the current across a heavy shared bus, instead of many thin individual feeds, removes 60% of the distribution loss. The I²R law is what makes the difference this large.

• The loss grows with the square of the current : For the same conductor, doubling the DC current quadruples the loss. Therefore rising RU power makes this problem worse faster than it makes the RU bill worse.
• A 4x to 10x increase in DC loss is expected : That is what happens if sites keep using per-RU #8 AWG trunk cables. Meanwhile RU power keeps growing.
• The bulk feed DC bus removes about 60% of the loss : In the worked example the loss per RRU at peak falls from 99.18 W to 39.51 W. The voltage at the load also improves, from 48.64 V to 51.87 V.
• It also simplifies the power plant : DC boosters are no longer required. Therefore conversion losses fall, complexity drops, and breaker positions are freed at the base of the tower.
• Aluminium is the right conductor here : For the same ampacity it costs about one third of copper and weighs about one half. On a tower, the weight saving matters as much as the cost.
• Do it during the 5G build : The tower is already being worked on. Retrofitting DC distribution as a standalone project pays for the climb twice.

How much energy does cooling consume, and what does liquid cooling change?

Cooling is the item I would not have expected to matter this much. It is the second largest block in the RAN breakdown, and the largest in the data centre. It reaches up to 40% of RAN energy consumption, with the exact share depending heavily on site configuration and location. In data centres it reaches 52% of total energy consumption. So this is a large target, especially in equipment rooms and centralised baseband sites where heat density is high.

The traditional coolant is air. The emerging alternatives use a dielectric liquid instead. A liquid has a much higher particle concentration than air, so its thermal conductivity is orders of magnitude larger. That is the whole basis of the improvement.

Two forms exist. In liquid immersion cooling the equipment is completely immersed in a thermally conductive dielectric liquid. Because the entire unit is in contact with the liquid, and not only a few components, immersion reaches the lower PUE of the two. Immersion itself comes in two variants. In the single-phase variant the liquid has a high boiling temperature and never changes state, and heat is exchanged through a heat exchanger. In the two-phase variant the liquid boils, so it changes from liquid to gas and holds the temperature at or just below the boiling point. That boiling point is typically around 50 degrees Celsius, and a condenser returns the gas to liquid. The two-phase variant generally achieves better energy efficiency, because the phase change carries more heat.

Direct Contact Liquid Cooling is the other form. Only selected components are liquid cooled, rather than the whole infrastructure. Therefore the achievable saving is lower than for immersion, and the resulting PUE is higher. However, DCLC is likely the nearer-term option. Migrating specific energy-intensive components such as CPUs, GPUs and accelerators is easier than immersing entire racks that contain equipment from several vendors.

The numbers give the scale. A legacy air-cooled cloud service provider data centre sits at a PUE between 1.3 and 2, depending on climate, ambient temperature, power delivery availability, backup arrangements and IT loading. One European operator's data centre in the northwest of France reports a PUE of 1.3. Immersion cooling has been shown to reach a PUE of 1.02 in similar scenarios, if all the IT equipment is immersed. A realistic mixed installation, where compute is immersed but storage and switches are not, achieves between 1.02 and 1.1. That still significantly outperforms air cooling. Separately, immersion cooling combined with energy reuse such as district heating has been shown to reduce carbon emissions by 45% against traditional data centre usage.

One caveat runs through all of this. Liquid cooling is being piloted in some RAN baseband units and IT data centres, but it is not yet mainstream in mobile networks. Getting the full benefit also requires the telecom equipment itself to be adapted. Moving parts such as fans and hard disks have to be removed, and the cooling loop has to be connected to a heat exchanger or a heat reuse system.

Figure 7. Immersion cooling reaches a PUE close to 1.0, so almost all the facility energy reaches the equipment. DCLC gives less, but it migrates far more easily.

• Cooling is up to 40% of RAN energy : It is 52% in data centres. Therefore cooling is second only to the base station equipment itself, and it is first in a centralised baseband site.
• Liquid beats air because of thermal conductivity : A dielectric liquid conducts heat orders of magnitude better than air, because of its far higher particle concentration.
• Immersion reaches PUE 1.02, air cooling sits at 1.3 to 2 : A realistic mixed installation lands between 1.02 and 1.1. The reason is that storage and switches usually stay out of the fluid.
• Two-phase generally beats single-phase : The liquid boils at around 50 degrees Celsius, and the phase change carries more heat than circulation alone.
• DCLC is the easier migration : It cools selected components such as CPUs, GPUs and accelerators. That is far simpler than immersing whole multi-vendor racks, although the PUE gain is smaller.
• The equipment has to change too : Fans and hard disks have to go, and the loop has to connect to a heat exchanger or a heat reuse system. Liquid cooling is not a drop-in replacement.

Does network disaggregation save energy, or spend it?

Disaggregation means running network functions as software on shared computing infrastructure. It sits in the long-term group, and it carries more caveats than anything else there. The reason is that it can cut energy and it can also raise it.

The saving comes from flexibility and scalability. With suitable management automation, the supply of network functions can be matched more closely to the actual demand. In addition, compute load can be moved to data centres that have access to renewable energy, and away from places that have less of it at that moment.

The cost comes from what is being replaced. A highly integrated network function runs custom software on custom hardware, and that combination is efficient by construction. Moving the same function to a general-purpose computing environment may well consume more energy for the same work. That is a near-term penalty against a longer-term structural gain.

There is also a measurement problem, and it blocks the business case rather than the technology. To know whether disaggregation helped, an operator has to determine which resources in the underlying compute infrastructure each network function used. Then the energy consumed by those resources has to be measured and reported back into network management. Until that chain exists, the saving cannot be quantified.

No percentage is published for this item, and further study is recorded as needed. That absence is itself informative, because every other row in the recommendations table carries a number.

• The saving is in matching supply to demand : Software functions can be scaled with load, and compute can be placed where renewable energy is available. Neither is possible with fixed appliances.
• General-purpose compute is less efficient per operation : Custom software on custom hardware is hard to beat on energy alone. So the first step of a migration can increase consumption.
• Attribution is the blocking problem : Energy has to be traced from a shared compute platform back to the individual network function. Without that, no saving can be proven.
• No figure is published, and that is deliberate : Disaggregation and cloudification are marked as needing further study, while every other solution in the table carries a quantified example.

What can advanced radio technologies add?

The last group changes the radio channel itself. The logic is that the channel decides how much power is needed to deliver a service. Modulation, required coverage, environment and geography all feed into that. Therefore a radio technology that transmits information more efficiently improves network energy efficiency directly. Two candidates are covered, and both are studied rather than deployed.

What does a reconfigurable intelligent surface offer?

A reconfigurable intelligent surface is a passive two-dimensional array. It is built from a large number of reconfigurable electromagnetic elements plus a control logic. By tuning those elements, the amplitude, phase and polarization of the reflected wave can be changed dynamically. Deployed on walls and ceilings indoors, or on buildings and signage outdoors, a RIS gives the operator a way to tune the wireless channel itself.

Field tests were run in three scenarios, and they measured RSRP and downlink throughput. The first scenario is outdoor-to-indoor. It uses a 4.9 GHz gNB with 100 MHz of bandwidth. The surface is 1 m by 1 m, with 32 by 32 equally spaced elements. There RSRP improved by up to 21 dB with a codebook approach. Downlink throughput improved by up to 4.4 times. The second scenario is outdoor, with a 3.5 GHz C-band gNB and 80 MHz of bandwidth. There RSRP improved by up to 10 dB, and throughput by up to 40%. That surface measured 40 cm by 40 cm. The base station to RIS link was line of sight at 60 m. The RIS to UE link was non line of sight at 52 m. The third scenario is indoor mmWave, with the direct path blocked by a building wall. There RSRP improved by up to 35 dB.

These measurements then fed system-level simulations of power consumption. The simulation result is the headline figure. Above roughly 5 bit/s/Hz of spectral efficiency, an optimally dimensioned RIS-assisted system is more energy efficient than the same system without RIS. The gain reaches up to 3.5 times.

Two qualifications belong with that number. The first is the word "optimally dimensioned", which is doing real work in the sentence. The second is that the throughput and RSRP figures are field measurements, while the energy efficiency figure is simulated on top of them.

What does distributed massive MIMO offer?

Distributed MIMO is a form of massive MIMO being considered for 6G. In DMIMO, all or at least a large number of the transmission and reception points in a wide coverage area cooperate. They jointly serve all the surrounding UEs on the same time or frequency resources. The TRPs act as one distributed massive MIMO array, coordinated by a central processing unit over a fronthaul network. The most ambitious form of this is cell-free massive MIMO.

The energy argument follows from two properties. First, DMIMO removes the traditional cell boundaries, and it significantly reduces inter-cell interference. Second, the distributed antennas provide a macro-diversity gain, so connectivity is more uniform than with co-located massive MIMO. Less interference and more uniform coverage together mean less transmit power for the same service.

The evaluation used system-level simulation in a dense urban enhanced mobile broadband environment, following the ITU evaluation guidelines. 84 users were served, with half concentrated around two hotspots and the rest uniformly distributed. Each system was simulated for a varying number of TRPs. DMIMO combined with adaptive carrier shutdown reached up to 2.5 times the network energy efficiency of cellular massive MIMO, measured in Mbit per joule. At the same time it delivered average per-user rates that cellular massive MIMO could not match.

Note the phrase "combined with adaptive carrier shutdown". The gain is not from the distributed architecture alone. It comes from the architecture plus an energy saving feature acting together, which is a useful reminder for the whole of this note.

• Both technologies work on the channel, not on the hardware : RIS reshapes the reflected wave, and DMIMO removes cell boundaries. A better channel needs fewer watts for the same service.
• RIS reaches up to 3.5x network energy efficiency : The gain appears above roughly 5 bit/s/Hz of spectral efficiency, and it assumes an optimally dimensioned surface.
• The RIS field measurements are strong on their own : RSRP improved by up to 21 dB outdoor-to-indoor at 4.9 GHz. It improved by up to 10 dB outdoors at 3.5 GHz. Indoors at mmWave, through a blocked path, it improved by up to 35 dB.
• DMIMO reaches up to 2.5x, with adaptive carrier shutdown : That is against cellular massive MIMO, in a dense urban eMBB simulation with 84 users. The saving feature is part of the result, not an extra on top of it.
• Simulation is not deployment : Both headline figures come from system-level simulation. In the RIS case the simulation is anchored on field measurements, which makes it the better evidenced of the two.
• DMIMO needs a fronthaul that most sites lack : Coordinating many TRPs from a central processing unit is a transport requirement as much as a radio one. That is part of why it sits in the long-term group.

What does the prioritised list of solutions look like?

A single table collects every quantified example in one place. I have reproduced it below in the original order, because the ordering is part of the message. Read it as a menu sorted by lead time, and not as a ranking by size of saving.

• The table is sorted by lead time, not by size : A 60% DC loss reduction sits below a 3% network saving in the list. That is because the 3% can start this quarter and the 60% cannot.
• One row has no number : Disaggregation and cloudification is marked "further study needed". That row is a measurement gap, and not a small saving.
• Two rows are ratios, not percentages : RIS at 3.5x and DMIMO at 2.5x are energy efficiency multipliers, expressed in delivered bits per joule. They are not reductions in consumption.
• Cooling appears twice, on purpose : DCLC gives less and arrives sooner. Immersion gives more and arrives later. Both are listed, because the choice depends on the site rather than on the technology.

How should these percentages be combined?

This section exists because the table above invites one specific mistake, and it is worth a minute of your time. The percentages are not shares of the same quantity. Therefore they cannot be added.

Consider what each one measures. The 3% and the 9% are shares of total network energy in the trial network. The 30% for a tri-band RRU is a share of what three single-band RRUs would have drawn. The 50% for the feeding path is a share of the feeding path loss only, which is itself a fraction of the transmitted power. The 60% for DC distribution is a share of the DC distribution loss, which is energy that never reached the radio in the first place. Adding these produces a number with no physical meaning.

The correct method is to carry each saving through the share chain. That calculation is worth following once, step by step. The RU is about 80% of base station equipment energy, and base station equipment is at least 50% of RAN energy. So the RU is about 40% of RAN energy. A 30% to 40% energy saving per RU therefore yields around 12% across the RAN. That is how a large per-unit percentage becomes a modest network percentage.

The same method sets expectations elsewhere. A 30% saving on the BBU is worth about 30% of 13% of 50%, which is under 2% of RAN energy. Therefore the BBU is a genuine target, and it is roughly six times less valuable than the RU. In the same way, the DC distribution saving is large as a percentage and small in absolute terms per RRU, at 523 kWh per year at peak load. It becomes material because a network has many RRUs, and because it costs almost nothing to operate once installed.

Three practical rules follow. Always state the baseline with the number. Always state what the percentage is a percentage of. Always check whether a reported figure is an energy reduction or a metric improvement, because the report contains both and they look identical in a slide.

• Percentages with different denominators cannot be added : One is 3% of a network. Another is 30% of what three RRUs used to draw. A third is 60% of a distribution loss. These are three different quantities that happen to share a symbol.
• Carry each saving through the share chain : A 30% saving per RU becomes about 12% of RAN energy, because the RU is about 40% of RAN energy. That multiplication is the whole method.
• The BBU is worth about one sixth of the RU : 30% of 13% of 50% is under 2% of RAN energy. Both are worth doing, but they are not worth doing first.
• Distinguish an energy saving from a metric improvement : 41% was an energy reduction per site. 260% was an improvement in the QoS-adjusted metric, and it combined lower energy with a doubled UE rate.
• Always name the baseline : 9% and 12% describe the same trial. One is measured against a rule-based network, and the other against a network with the features switched off.

Where did this work go next?

A 2023 snapshot ages quickly in this area, so I went looking for what came after it. The Phase 2 report was approved in August 2023 and published in October 2023. It closes with a recommendation rather than a conclusion. Operators are advised to review the saving potential of each option, and further study is requested on the saving potential of machine learning, of network disaggregation and of cloudification. Additional study is also requested on the state of the art in energy management and in renewable energy solutions.

Renewable energy needs one careful distinction, and it is an easy one to lose. Deploying solar or wind at a RAN site does not save energy. It changes where the energy comes from. What it does reduce is carbon emissions, energy bills and dependence on the grid. Advances in batteries, fuel cells and energy management systems improve that picture further.

The programme continued. A third phase followed, and its results were collected in a roadmap document published in July 2024. That roadmap keeps the same three groups and the same time-horizon structure, so the framework in this note still applies. What changes is the emphasis. Measurement moves to the front, with static and dynamic laboratory procedures defined for assessing base station equipment. AI moves from a modelling aid to a control function, used to predict energy efficiency and load states and to identify low-efficiency sites. Coordination between network elements becomes a topic in its own right, covering multi-carrier scheduling alignment, spectrum sharing combined with carrier shutdown, and RAN sharing between operators. Interworking between the mobile network and the power supply also appears as a new area.

The direction of travel is consistent across both phases. The early gains come from configuring what is already installed. The later gains come from coordinating equipment that was previously optimised in isolation.

• Renewable energy is a carbon lever, not an energy lever : It does not reduce consumption. It reduces emissions, bills and grid dependence, which are different objectives with different metrics.
• Measurement was the first gap to be closed : The following phase defines static and dynamic laboratory procedures for base station equipment. Without a repeatable measurement, a comparison between vendors is not possible.
• AI moved from modelling to control : In Phase 2 it fits RU power models and tunes thresholds. In the roadmap it predicts load and energy efficiency states, and it identifies which sites to act on.
• Coordination is the next structural theme : Multi-carrier scheduling alignment, spectrum sharing with carrier shutdown, and inter-operator RAN sharing all save energy by removing duplicated readiness.
• The framework did not change : Process optimisation, engineering optimisation and new technologies still map onto short, medium and long term. So the way of reading the options in this note carries forward.

Reference

The two documents below were read in full for this note. The list after them collects the external sources those documents rely on for the figures quoted here, so that a reader can go back to the original measurement.

• NGMN Network Energy Efficiency Phase 2 (v1.0, 10 October 2023) : NGMN Alliance, Green Future Networks - the primary source for this note. Landing page at ngmn.org
• Green Future Networks: A Roadmap to Energy Efficient Mobile Networks (v1.0, 2 July 2024) : NGMN Alliance - the Phase 3 follow-on used for the last section. Landing page at ngmn.org

Sources cited by the reports above for the figures used in this note:

• NGMN Green Future Networks: Network Energy Efficiency (15 December 2021) : NGMN Alliance - the Phase 1 report, and the source of the RAN and base station energy breakdown
• 3GPP TR 38.864 : Study on Network Energy Savings for NR - the base station power consumption model used for evaluation
• RP-223540 : 3GPP Work Item, Network energy savings for NR
• arXiv:2212.04318 : Power Consumption Modeling of 5G Multi-Carrier Base Stations, A Machine Learning Approach - the ML framework behind the fitted RU power model
• Nature Electronics, 2020 : Energy-efficient 5G for a greener future - the source of the 4077 W and 950 W radio unit comparison
• ETSI TS 103 586 : Environmental Engineering (EE); Liquid cooling solutions for ICT infrastructure equipment
• IEEE Communications Magazine, vol. 58, no. 1, 2020 : Towards Smart and Reconfigurable Environment, Intelligent Reflecting Surface Aided Wireless Network - the RIS reference used by the report
• IEEE Transactions on Wireless Communications, 2017 : Cell-free massive MIMO versus small cells - the cell-free reference used by the report