Taming the energy use of gaming computers
- First Online:
- Cite this article as:
- Mills, N. & Mills, E. Energy Efficiency (2016) 9: 321. doi:10.1007/s12053-015-9371-1
- 210 Views
One billion people around the world engage in some form of digital gaming. Gaming is the most energy-intensive use of personal computers, and the high-performance “racecar” systems built expressly for gaming are the fastest growing type of gaming platform. Large performance-normalized variations in nameplate power ratings for gaming computer components available on today’s market indicate significant potential for energy savings: central processing units vary by 4.3-fold, graphics processing units 5.8-fold, power supply units 1.3-fold, motherboards 5.0-fold, and random access memory (RAM) 139.2-fold. Measured performance of displays varies by 11.5-fold. However, underlying the importance of empirical data, we find that measured peak power requirements are considerably lower than nameplate for most components tested, and by about 50 % for complete systems. Based on actual measurements of five gaming PCs with progressively more efficient component configurations, we estimate the typical gaming computer (including display) to use approximately 1400 kWh/year, which is equivalent to the energy use of ten game consoles, six standard PCs, or three refrigerators. The more intensive user segments could easily consume double this central estimate. While gaming PCs represent only 2.5 % of the global installed PC equipment base, our initial scoping estimate suggests that gaming PCs consumed 75 TWh/year ($10 billion) of electricity globally in 2012 or approximately 20 % of total PC, notebook, and console energy usage. Based on projected changes in the installed base, we estimate that consumption will more than double by the year 2020 if the current rate of equipment sales is unabated and efficiencies are not improved. Although they will represent only 10 % of the installed base of gaming platforms in 2020, relatively high unit energy consumption and high hours of use will result in gaming computers being responsible for 40 % of gaming energy use. Savings of more than 75 % can be achieved via premium efficiency components applied at the time of manufacture or via retrofit, while improving reliability and performance (nearly a doubling of performance per unit of energy). This corresponds to a potential savings of approximately 120 TWh/year or $18 billion/year globally by 2020. A consumer decision-making environment largely devoid of energy information and incentives suggests a need for targeted energy efficiency programs and policies in capturing these benefits.
KeywordsInformation technologiesComputing energy useGaming computers
Estimates placed the flow of digital media to US households at 6.9 zettabytes (ZB; 1021 bytes) per year in 2012, of which 2.5 ZB (34 %) was attributed to gaming (Short 2013). US households are projected to spend 211 billion hours of gaming in 2015, more than the time spent on the telephone, mobile computing, or messaging. Use has doubled since 2008. The 43.6 million “extreme” and “avid” gamers spend 4.4 h/day in the activity (all platform types) versus 7.2 h/day for the 10 million “extreme” gamer subgroup (Short 2013).
The global count of people utilizing gaming computers was estimated at 54 million in 2012 (33 countries studied) and projected to grow to 72 million together with sales of related computer hardware of $32 billion by 2015 (Business Wire 2012). About half of the 100 million PCs with discrete graphical processing units (GPUs) shipped in 2014 were purchased by consumers, with the other half destined for workplace environments (Peddie 2014).
Computer gaming is engaging an increasingly diverse user base. These consumers spent $22 billion on gaming software in 2013 (ESA 2014), with the global market estimated at $100 billion (Brightman 2013). The scale and growth of this activity calls for assessment of the associated energy use.
Just over half of all US households own a game console, with the average player being 31 years old and with males and females engaged in roughly equal proportions. Previous studies exploring the energy implications of game console use found average unit electricity use to be 102 kWh/year for the installed US stock (excluding the connected display) and 64 kWh/year for new sales as of mid-2012 (Webb et al. 2013).1 There is ongoing debate about game console utilization, with recent studies finding that this may have been previously overstated (Desroches et al. 2015).
We found no prior studies focusing on the aggregate energy used by gaming computers. One assessment (Ecova 2012) examined the idle power demand of graphic processing units embedded in gaming computers, and another (Brocklehurst and Wood 2014) explored whether these machines would be able to meet the ENERGY STAR v6.0 requirements, based on pooling diverse test results from third-party sources (not standardized for factors such as choice of motherboard, duration of sleep mode, overclocking, operating system, software running during testing, etc.). Their results were confounded by differences in test procedures.
This article provides new information based on nameplate performance of gaming computers and their components together with direct measurements. Efficiency opportunities are identified. Using measured data, we produce the first global estimate of the associated current and projected energy consumption and savings potential.
Components, architecture, and efficiency options
Gaming computers contain the same generic components as conventional computers. However, the performance requirements of these machines entail far higher energy intensities, and in many cases, multiple components (e.g., GPUs, hard drives, displays) are used. Protocols for benchmarking the computational performance of gaming computers involve running a preset gaming process and collecting metrics. Some benchmarks focus on central processor performance (e.g., Cinebench); others focus on the graphics (e.g., Unigine Heaven; see http://www.maxon.net/products/cinebench/overview.html and https://unigine.com/products/heaven/). Component product literature, however, emphasizes nameplate estimates of power requirements, rather than actual performance or power needs under a given mode of operation. As discussed below, accurate energy use calculations cannot be made with nameplate data. However, no standardized test procedures exist for evaluating gaming actual computer energy use, which perpetuates market reliance on over-estimates of nameplate data.
Components of gaming computers and efficiency opportunities
Energy saving strategies
30 to 150 W
13–65 W/GHz of max supported CPU
More efficient capacitors; improved power delivery efficiency and control. Some motherboards allow the user to disable components not in use (e.g., HDMI, PCI-E slots, or SATA ports).
Central processing unit (CPU) architecture
37 to 220 W
Decreased size and increased transistors per unit area (less leakage). Power scaling (e.g., Intel Sandy Bridge (85 W) vs. Ivy Bridge (77 W) vs. Haswell (65 W) illustrate the generational progression). C-state (aka “C-mode”) capabilities enable CPU to vary power draw as a function of workload, with particular emphasis on increasingly sophisticated sleep modes. There are currently 13 C-state gradations, some of which can be changed by the user in the Basic Input/Output System (BIOS). Selected voltages can be reduced within the CPU without reducing performance (but with reduced stability CPUs can be underclocked to reduce power consumption (but with reduced performance). Multiple cores may or may not affect efficiencies, depending on computational activity and software.
Graphical processing unit (GPU)
75 to 500 W
Decreased size and increased transistors per unit area (less leakage). Power scaling (e.g., NVIDIA Fermi vs. Kepler vs. Maxwell). GPUs can be underclocked for additional energy savings (but with lower performance). Modes exist for disabling GPUs when the display is off. Displays with “anti-tearing” features enable use of lower-power GPUs.
Low single-digit watts, but can be many fans (typically 5–6) in a single computer
Efficiency of air movement. Automated power-down at low loads. Improved blade designs. Reduced fan count commensurate with efficiency improvements elsewhere in the system.
DDR (2.5 V) ≥ DDR2 1.8 V) ≥ DDR3 (1.5–1.65 V) ≥ DDR4 (1.2–1.35 V)
Reduced voltages. Fewer higher-capacity modules (“sticks”).
HD (~10 WW) ≥ SATA SSD (~5 W) ≥ PCI-E SSD (~3 W)
Switch from mechanical to solid state with significant performance boost in reads and writes.
Power supply unit (PSU)
Intrinsic energy use only from dedicated fans. Indirectly associated with losses due to power conversions for downstream loads.
70 % efficiency ≥ 80 % (80Plus threshold) ≥ 94 % (80Plus Titanium; all at 50 % load)
Efficiency; some units are fan-less, saving several watts; others curtail fan use until high power thresholds are reached. Sizing to match load is important for peak efficiency, although less so as the industry has attained more consistent efficiencies across the load range.
15 to 77 W (23–34 in. size range)
Technology choice (CRT vs. LCD/LED, +backlighting strategy, as well as techniques to avoid image tearing with lower GPU speeds. Power management (e.g., sleep mode), dynamic dimming as a function of room light levels, and occupancy-sensor-initiated sleep mode. Improving transmissivity of film stack to improve luminous efficacy. Display-specific PSUs also present efficiency opportunities.
Various energy management tools are available via the OS.
Tuning voltages to required performance level. Constant voltage vs. ASUS EPU engine.
Curtailing operation of some or all components after designated time. Monitor sleep functionality; GPU staged control where unit has multiple processors (e.g., AMD “zero-core” technology) or thermostatically controlled fans.
Intelligent automatic fan control
Variable speed control as function of eight internal temperature sensor signals. Some GPUs allow user to specify desired fan speeds as a function of temperature. T-Balancer: Big NG.
Central processing unit
Graphics processing unit
Memory and storage
Gaming computers require dedicated cooling systems in order to avoid overheating, even at idle. Active cooling is typically provided to each power supply unit (PSU), CPU, GPU, and motherboard as well as to the general environment within the computer chassis. In a CPU air cooler, there are typically one to three fans driving hot exhaust air across a heat sink. With liquid cooling, a heat exchanger mounts to a particular component (CPU, GPU, motherboard, or memory) and directs the coolant over a heat-exchange plate that is in direct contact with the component. Liquid cooling is often preferred because it allows the processor to achieve higher overclocks (enhancing computational performance at lower temperatures). We measured CPUs with and without liquid cooling, and no change in energy use was observed.
Power supply units
While typically not hardwired to the gaming computer itself, with the exception of notebooks and consoles, displays are integral and energy-intensive elements of the system. Moreover, although independently powered, display choice influences power requirements and performance of the GPU in gaming mode. Energy use varies widely as a function of technology, screen size, and resolution. The dramatic technology transitions that have occurred in displays, resulting in significant energy benefits, have been driven more by the desirable form factors and image quality than by energy savings. Countervailing trends are the transition from VGA/SVGA to HD/1080p, to 4 K displays, as well as the use of multiple displays. The net effect is that GPUs must drive many more pixels than was the case just a decade ago.
Gamers have historically been irked by visual anomalies such as image “tearing” and “stuttering”. Tearing occurs when a frame is outputted by the GPU when the monitor is in the middle of a refresh. One solution to this issue involves enabling V-Sync (Vertical Sync) where tearing is eliminated by forcing the GPU to wait until the monitor is ready to refresh the next frame. This can cause unacceptable delays in screen refreshes, i.e., stuttering. New technologies such as G-sync (NVIDIA, hardware) and FreeSync (AMD, software) allow more effective communication between the GPU and the monitor. When these run during gameplay, the GPU tells the monitor when to refresh, resulting in little to no stuttering and no tearing. If the frame-rate in the game is low, these approaches will synchronize the GPU output with the game’s capacity to render. This saves energy since, even at around 30 to 50 frames per second (FPS), the gaming experience becomes smoother to the gamer’s eye, enabling the gamer to specify a GPU with lower nominal performance (and power requirements). With these technologies, manufacturers claim that gaming will be as smooth as with a higher-power GPU.
Nameplate power estimates and energy use of gaming computers
The capabilities and performance of gaming computers vary widely, depending on which components are selected. Components with similar computing performance must be compared in order to evaluate baseline energy use and savings potential in a meaningful way. While many other consumer products (including game consoles) are typically evaluated in terms of total system load, gaming computers can also be evaluated at the component level. However, it must be kept in mind that nameplate power values are often far higher than maximum power use.
The resulting scenarios for high-power, typical-power, and low-power configurations nominally draw 923, 601, and 331 W, respectively (including displays). Note that in many warm locations, or in many large commercial buildings, significant additional electricity use would be required for air conditioning (not accounted for here) needed to remove the heat produced by these machines. In other locations the computer’s waste heat may be useful for part of the year.
Individual gaming computers could have higher power consumption than these reference machines. This can arise not only where less efficient components are used but also where multiple monitors, GPUs, or storage devices are employed. Additional discretionary energy-using components (internal or external) include sound cards, digital-analog converters (DACs), headphones, amplifiers, speakers, networking equipment, RAID cards, powered keyboards, pointing devices, and decorative lighting. The most energy-intensive component in the gaming computer is the graphics processing unit (GPU), and 1.4 graphics cards were sold for each computer sold in 2014 (JPR 2014)2; only one GPU is assumed in these reference machines. Overclocking also increases power consumption and waste heat, as does disabling power management features.
While on the one hand, the above-referenced market data suggest exceptionally high energy use, it is also important to observe the large variation in the various intensity metrics. The history of computing has shown sustained and significant strides in intrinsic energy efficiency (e.g., calculations per second per watt) and that is evident in the gaming PC arena where efficiencies double every 18 months (Koomey et al. 2011). That said, consumer demand for increased performance has risen even more quickly, with the net effect of rising absolute energy use. These points notwithstanding, given the limitations of nameplate information it is important to explore the actual outcomes by examining measured data.
Measured power and energy benchmarks
Extending nameplate power to estimates of actual energy use is not straightforward. The resultant energy use depends on differences between actual and nameplate capacity as well as the mix of usage modes and duration of use in each mode (e.g., off, sleep, idle, active gaming, video/movie playback, and Web browsing). For example, Webb et al. (2013) found that approximately half of the on-time for game consoles is in “gameplay” mode. Each game or process (e.g., 3D rendering) has its own energy intensity. Moreover, there are a variety of levels of computing demand even within the general activity of “gaming,” and energy use is also software specific.
Little measured data has been collected for gaming PCs and their sub-components. The performance of a given component relative to that of other components in the system will also vary significantly depending on the mode of operation. In one example, a particular motherboard ranked average compared with 11 others (using identical CPU) when in long-idle mode, above average in idle mode, and lower than average in active computing mode (Cutress 2014).
We constructed a baseline gaming computer using popular components on the US marketplace as of December 2014. We then measured power requirements and energy use by mode while running common gaming performance benchmark software. Our test-bench machine contains a motherboard that utilizes the X79 (aka Patsburg) chipset and an LGA 2011 CPU, noted by others (Brocklehurst and Wood 2014) to be among one of the highest performance Intel platforms on the market. (As of August 2014, X79 was succeeded by the X99 (Wellsburg) chipset and LGA 2011–3 Socket).
We performed a range of system-level measurements in different modes of operation, capturing loads from “off” to full gaming mode status. We adopted estimates by Short (2013) for average times spent by US gaming computer users in various modes of operation. We included short idle times (measurements over the interval of 5 to 10 min after cessation of user inputs) as well as long idle times (after idle for 10 min of idle) per the ENERGY STAR v6.0 test procedure and no “2D” operation (only benchmarking software was running during tests) (USEPA 2013). Established software performance benchmarking tools were utilized to stress test the components and create replicable results under conditions used more broadly in the industry. One-second power data were taken with Watts-Up Pro ES data logger. Internal and after-market software enabled sub-metering in some cases (PSUs, CPUs, and GPUs).
Measured power consumption and energy use for our base case varied significantly as a function of usage mode. Measured peak electricity demand in active gaming mode at 512 W is six times that of a typical desktop computer and its associated display and three times that in idle mode (Urban et al. 2014). The mode-weighted-average power draw during on-time was 212 W.
Disparities between nameplate and actual component power requirements
Nameplate rating (W)
Measured power (W, at peak)
CPU: Intel Core i7 4820 K (at 3.7 GHz, rated)a
CPU: Intel Core i7 4820 K (at 4.5 GHz, overclocked)a
CPU: Intel Pentium G3258 (at 3.2 GHz, rated)a
CPU: Intel Pentium G3258 (at 4.0 GHz, overclocked)a
GPU: NVIDIA Geforce GTX 970 (at 1102 MHz, rated)b
Apple HD Cinema Display
Apple Thunderbolt Display
Full-base system benchmark: CPU testa
Full-base system benchmark: gaming modec
We document differences in nameplate and measured power values in Table 2. This effect is compounded where multiple components are evaluated when assembled as a system, with a 49 % disparity during gaming mode in the case of our built-up system. One important ramification of these disparities is the degree to which PSUs will likely be oversized if nameplate performance is relied upon.
The results indicate unit energy consumption of 1394 kWh/year (based on an average of 4.4 h/day in gaming mode), including the display. The “Avid” user sub-segment (29.5 million people, USA) spends 3.6 h/day gaming, uses 1300 kWh/year, while the “Extreme” user segment (8.1 million people) spends 7.2 h/day uses 1890 kWh/year (36 % more; utilization rates from Short 2013). For the typical gamer (4.4 h/day, weighted average of Avid and Extreme), we found that a much larger proportion of total energy (80 %) occurs in modes above idle than is the case for traditional personal computers, which have low computing loads (Beck et al. 2012). High-performance computers in work environments (not included in this analysis) will also have high consumption where there are more average daily hours of use.
The cost of this electricity would be on the order of $200/year at typical household electricity prices (and easily $500/year where tariffs are usage dependent, e.g., with an inverted-block design). This, in turn, corresponds to emissions of approximately 1700 lbs (780 kg) of carbon dioxide/year at US-average electricity emissions factors (USEPA 2010).
These estimates are likely conservative, as we assume only one display per user, no peripherals such as audio equipment, and no overclocking of CPUs or GPUs, and “Power saver” settings in the operating system.
Energy efficiency potential
To explore the potential for efficiency improvements and corresponding energy savings, we made a series of progressive hardware improvements to the system and measured the response. These included a more efficient PSU, GPU, CPU, motherboard, and display.
Additional savings can be achieved through operational settings. One analysis based on adjustments to the CPU and motherboard achieved 27 % savings in standby power, and 26 to 30 % savings in active mode (3DMark and Cinebench benchmarks, respectively) without a reduction in performance (Crijns 2014). Additional adjustments involving underclocking and voltage management yielded 44 and 64 % allowing for 16 and 30 % reductions in performance under the same benchmarks. Combined with the efficiency gains achievable with improved CPU, GPU, and motherboards can thus be expected to yield a total of more than 75 % annual energy savings.
Some efficiency improvements have ancillary benefits. For example, the base GPU in our comparison experienced internal temperatures of 91 °C during the Unigine Heaven benchmark trial, which fell to 65 °C with the more efficient unit due to improved cooling, power delivery, and power consumption. This supports increased reliability and service life, while reducing fan speeds and noise and achieving lower temperature environments for nearby components.
Role of consumer information environment, decision-making and behavior
Gaming computer purchasers face many barriers to making energy-efficient choices. Most components bear no energy-related information on their packaging or when bought on-line without packaging. This includes the most energy-intensive components (graphics cards and CPUs), which do not even carry nameplate power estimates on their packaging or on the product itself. Even spec sheets do not always contain this information. Integrated systems also typically lack information on requirements, aside from the nameplate power of typically oversized PSUs.
Thus far, no labeling programs differentiate the energy performance of gaming computers. The highest long- and short-idle power requirement among ENERGY STAR-rated desktop computers are 33 and 63 W, respectively, which suggests that no gaming computers have received ENERGY STAR ratings. At least in the USA, mandatory energy efficiency standards do not exist for any components found in gaming computers.
Retail salespeople are poorly equipped to coach buyers. Some that we interviewed use highly imprecise rules of thumb when recommending power supplies, e.g., based on unreliable nameplate performance of the associated graphics card plus a “safety margin.” It is encouraging that some industry watchers have proposed that metrics be developed to consider total cost of ownership (including energy costs) (Pollak 2010), but this has yet to become mainstream thinking.
Power supplies have received more attention over the past decade than other gaming computer components, leading to the voluntary 80Plus program (Calwell and Ostendorp 2005). The program includes a staged rating system denoted by bronze, gold, platinum, and titanium. In retail environments, we observed misleading product labeling, where words like “gold” and “silver” were used in a way that masks the absence of an actual 80Plus rating.
Aside from 80Plus, energy test procedures are not standardized, creating considerable confusion in the consumer information environment. For example, three Websites rate an identical motherboard at 62, 92, and 98 W (a 58 % difference across the range)—all at idle and independent of associated CPU (see http://www.guru3d.com/articles_pages/asus_z97_sabertooth_mark_1_motherboard_review,8.html; http://www.kitguru.net/components/motherboard/luke-hill/asus-sabertooth-z97-mark-1-motherboard-review/12/; http://www.tweaktown.com/reviews/6345/asus-sabertooth-z97-mark-1-intel-z97-motherboard-review/index8.html). Such differences could arise from a range of factors not typically standardized (or even disclosed) in test reports. Examples include disparate power supplies or power management. Standardized test procedures are clearly needed.
Technical efficiency ratings reach only so far, as user behavior is an over-riding factor in ultimate energy use. As noted previously, hours of use vary widely, as do consumer desires regarding extreme performance capabilities, display count and area, peripherals, etc. The sports-car analogy applies here in that technical energy savings are easily “taken back” in return for increased performance and corresponding energy use.
The net “worst-case” effect of consumer-determined factors is the high-power multi-display system depicted in Fig. 1. For perspective, that system entails three-times the nameplate power of our “typical-power” case and seven times that of the “low-power” case shown in Figs. 9 and 10.
Global energy use
Global gaming computer energy use in context: 2012
Gaming PCs: pre-builta
Gaming PCs: user assembled
Gaming PCs as fraction of total (%)
Unit energy consumption (kWh/year)b
Installed base in 2012 (million units)
Total energy consumption in 2012 (TWh/year)
We find that, although they represent only 7 % of PC, notebook, and console gaming platforms, gaming computers were responsible for electricity use of 75 TWh/year in 2012 (or approximately $10 billion/year) equal to 30 % of all energy use across this array of devices. Placed in a broader context, this represents about 20 % of electricity used by all PCs, notebooks, consoles, and tablets (Table 3).
As noted previously, users with multiple displays, multiple graphic cards, or other discretionary components will require even more energy. Additional energy will also be used in association with air conditioning in hot climates. Trends in technology and behavior (hours of use, by mode) may prove to be as important determinants of energy demand as changes in the hardware itself. Prior macro-level studies have not isolated the energy use by these machines from that of conventional computers.
The potential to reduce energy demand from gaming computers by more than 75 % is enhanced by the very rapid turnover of equipment (several years at the most), the ability for individuals to specify high-efficiency components (new or retrofit), and the significant co-benefits of energy efficiency enhancements for equipment performance, thermal management, and reliability. One of the more pronounced historical examples of technological process is the simultaneous 10-fold improvement in speed of RAM, accompanied by a 13-fold reduction in power requirements (Fig. 6). A key illustration of current opportunities are fan-less PSUs, which not only save significant energy due to the high efficiency associated with eliminating the need for cooling but also trim approximately four constant watts of base load demand, while attaining reduced noise and increased reliability by eliminating the dedicated fan altogether.
There is a wide range of energy use among individual gaming computer components as well as integrated systems. The metrics we computed suggest a correspondingly wide range in efficiencies, i.e., energy use for a given level of computing performance. This demonstrates that high performance can be attained without compromising efficiency. The energy use of gaming computers is significant, and growing, and projected to more than double by 2020 assuming today’s efficiencies and current projections of an increasing installed base of equipment. Overall efficiency improvements of 75 % or more are attainable, which would translate to savings of approximately 120 TWh/year or $18 billion/year at a global scale in the year 2020. Assumptions underlying the typical computer modeled here likely understate energy use in practice.
The results of prior studies have been confounded by uncertainties introduced by relying on nameplate rather than measured data, as well as disparate test conditions and test procedures. We find that nameplate power estimates for the key components in gaming computers significantly exceed power use in practice (on the order of 50 %) and their direct use can thus yield overestimates of energy use. This problem requires attention through further testing under as-used conditions and applied towards improved consumer information and ratings. The energy requirements of specific gaming applications can also be evaluated.
From a technological standpoint, component efficiencies will no doubt continue to improve. Advanced control strategies are also important. Unlike almost all other energy-using products (including commodity PCs), a large share (one third) of gaming computers are specified and assembled by end users. This opens up a unique opportunity for interested consumers to attain efficiencies otherwise unavailable on the market. There is a promising trend towards more efficient notebook-format gaming computers. This has historically been difficult given the relatively large physical dimensions and weight of high-performance components and severe challenges in thermal management and battery life within the small form factor of notebook computers. Gaming notebooks, however, do not commonly deliver the same computing performance as do desktops but are improving.
Our macro-level results are certainly preliminary in nature, and suggest that the issue calls for much more rigorous analysis, which, in turn, requires the collection of more market data. In the future, finer-grain data on equipment stocks, energy using characteristics, and user behavior will allow for more precise and disaggregated energy-use estimates (e.g., in homes versus workplaces, the latter of which is not incorporated in our analysis). The additional gaming-related energy use of general-purpose computing devices also remains to be estimated. To enable improved energy analyses as well as better consumer decision making, standardized methodologies should be developed to more rigorously and consistently benchmark and normalize energy use and peak power demand of computers as well as that for specific games.
The mainstream gaming computer industry does not emphasize energy use or efficiency, consumers do not have ready access to the information needed in order to make informed decisions, and energy analysts and policy makers have only begun to identify the importance of this particular energy end use. Policies proposed for addressing other types of household electronics (OECD/IEA 2009) and game consoles in particular (Webb et al. 2013) could be beneficially applied to gaming computers as well. More vigorous energy programs and policies are needed to mitigate the energy consequences of the very fast-growing worldwide market for gaming computers.
It is important to consider learning-curve effects. Console launch models are typically two or more times as energy intensive than the given model’s stabilized performance once several generations of design refinements have been made (Delforge and Horowitz 2014); for example, the 2006 release version of PlayStation 3 required 180 W in “game play” mode, which ultimately stabilized at 70 W in the 2013 version.
This industry-wide statistic includes all types of desktop computers, while virtually all machines incorporating multiple graphics cards are gaming PCs (which are a small segment of the overall market). Thus, this value is likely a conservative reflection of the actual practice. Having multiple graphics cards is a very widespread practice among gamers, and some machines are even shipped from the factory with two installed.
We thank Jon Green, Oliver Kettner, Jon Koomey, Bruce Nordman, Ted Pollak, Brian Strupp, and three anonymous reviewers for their support and constructive comments.