Our motto is “Gaming Energy Efficiency without Performance Compromise”. This work is
sponsored by the California Energy Commission, and covers the full spectrum of
non-battery-charged gaming platforms, as well as gaming applications.
Meet our team and find out more about our project here.
Phil Eisler, General Manager of NVIDIA’s GeForce NOW Cloud Gaming service, discusses high-performance gaming in the cloud
For a given gaming PC, power in active gaming mode varies by a factor of three across the simulated benchmarks and a factor of two across individual real-world game titles
Online gaming is eating more and more bandwidth
Measured system power varies by seven-fold across representative PCs and four-fold across laptops
Notable Industry Activities
New data and analysis from the European Commission’s Efficient Gaming initiative shows great technological progress, but also challenges in establishing useful benchmarks
Relaxing virtual reality rendering loads at the periphery of vision saves significant amounts of energy
Six generations of desktop GPUs: Efficiency increasing, while power use remains relatively constant
Our new report offers a major assessment of the “An Energy-focused Profile of the Video Gaming Marketplace”
Comings & Goings
California Energy Commission (CEC) meeting of plug loads researchers
GPU Technology Conference
Green-up Your Game
Tuning up your rig
Buying new gear
Green Gaming News interviewed Phil Eisler, General Manager of NVIDIA’s GeForce NOW Cloud Gaming service. We talked with Phil about the energy considerations of emerging server-side gaming in comparison with the traditional client-side approach.
PE: Definitely—we see this trend as bringing more people to the world of gaming. The main goal of GeForce NOW is to make it easy to enjoy great games. Today, there are many barriers to PC gaming; like buying a desktop PC, installing new GPU upgrades, downloading new drivers, downloading games and installing patches. While this experience is worthwhile for some gamers, there are many more that want a simpler option, which is where GeForce NOW comes in. All gamers have to do is click on the GeForce NOW app on any Mac – and soon on PCs – and they get streaming access to a GeForce gaming PC in the cloud. This machine is ready with the latest hardware and software, and games are ready to play in seconds. Making PC gaming more accessible will expand the market.
Today, we’re streaming at 1080p resolution and 60 FPS at 40 Mbps, which matches the local playing experience for most users. If you’re playing at 4K or 120 FPS, then you’ll want to stay with your local GeForce GTX PC. The network adds about 30ms of latency that isn’t noticed by most users, but competitive gamers will still prefer a local GeForce GTX PC.
GGN: Presumably in the near term, existing gamers would use their current rigs with your service. Approximately what electric load would remain on their local CPUs and GPUs compared to when they are gaming solely on their own machines?
PE: Less than 15W is required to decode a GeForce NOW stream locally when a lower-powered device like SHIELD TV is being used on the client side. Regular PCs will use more depending on their non-active power requirements.
GGN: What is the power rating of the GPUs being used in your data centers?
PE: We use two types of GPUs in our data center to offer gamers different options. They range from 75 to 150W of electrical power. We use Tesla P40s in our data centers that have a theoretical maximum TDP of 250W.
GGN: What is the CPU load allocation method within the server?
PE: We assign a dedicated virtual gaming PC for each user. Each user is assigned up to 4-Core/8-Thread of CPU capacity, which is the same as the Intel Core i7 processors found in most desktop gaming PCs. Most games are designed for 4C/8T and run smoothly on our virtual machines.
GGN: What is your estimate of relative energy use for a given GeForce NOW gaming session as well as standby time (server plus associated facility infrastructure and downstream network) compared to what that would be the case in a strictly client-side session for the same game title?
PE: We haven’t done an analysis with this comparison yet. We always use the newest and most efficient GPUs, so we are more efficient than older PCs. Also, we can save energy for gamers who typically leave their gaming PC powered on to get updates since we do that for you on the server side.
GGN: Do you see opportunities for managing server-side energy use in the future such that less power will be used overall compared to strictly client-side gaming?
PE: NVIDIA makes each new generation of GPU more efficient than the previous generation. GeForce NOW will encourage earlier adoption of the newer, more efficient GPUs. We also continue to investigate ways that we can save power in our data centers. For example, we can suggest the right size of virtual gaming PC for your selected game and screen size, which can save the gamer from using excess power with an over-spec PC.
GGN: Assuming you're co-locating in others' data centers, are you involved in the infrastructure choices beyond the servers themselves, such as space conditioning and air management?
PE: Our first data centers are co-location. However, we’re looking at bigger data centers where we can apply NVIDIA engineering and technology to drive industry-leading PUE ratios. We’re looking beyond the servers including the data center HVAC systems for more efficiency in the future.
Defining the energy use of a gaming PC is no easy task. Among the many variables are which particular game (or benchmark) is being run. We ran four of our nine gaming rigs, selected to bracket the range of products available in the market, through some paces. We ran 10 different human-played games and 11 different benchmarks on each rig with surprising results. System power in active mode varied by a factor of three across the simulated benchmarks and by a factor of two across real-world game titles. Power also varied significantly for a given system, depending on which game or benchmark was being used. The absolute energy results for the benchmarks varied from about 125W to 375W on the most powerful of the machines tested. The results for the same PC varied from about 125W to 225W for actual gameplay. The relative rankings of the results remained mostly consistent across machines, although not in every case.
GREEN GAMING NEWS
Issue Number 4 - October 23, 2017
Note: Energy use is that of the computer only, excluding display in the case of desktop units. Subsequent testing showed that the H1 machine was by far the highest energy-using machine in our line-up. (Note: Figures updated May 30, 2018)
Clearly, the choices of hardware as well as software have significant implications for energy use. Power requirements are of course just half the equation. User experience is the other side. Watch this space for our related assessments in the coming months.
Gaming is rapidly expanding into the cloud [see Trendsetter Interview in this issue], thereby extending the associated energy use into data centers. No detailed information is yet available on the relative allocation of energy use between the local client and the network of supporting core and edge data centers.
According to Entertainment Software Association surveys, 51% of the most frequent gamers play cloud-based games at least once weekly, for an average of 0.9 hours per day. According to Statistica, among college students, approximately 35% play daily and 75% play at least weekly.
Nielsen data suggests that the shift towards online gaming is strong, with 21% of 7th-generation console hours spent in that mode in 2010, increasing to 28% for 8th-generation consoles in 2014. Console players now spend more time playing online games than offline games. Statistica placed the value at 46 million console gamers in the U.S. as of 2014.
An entire gaming genre—Massively Multiplayer Online Role Playing Games (MMORPG)—is intrinsically cloud-based. Statista reports that one third of all US gamers engage in MMORPG gaming, while NPD finds that 70% of the 34 million “core gamers” do so. As far back as 2012, PC gamers reported spending 34% of their total gaming time in on-line mode, according to PwC.
On-line gaming is projected to be the fastest-growing segment of residential Internet service, with the 1.1 billion users in 2015 growing to 1.4 billion by 2020.
Cisco has noted that “if cloud gaming becomes popular, gaming could quickly become one of the largest Internet traffic categories”. If achieved, virtual-reality streaming could become an important additional driver. Cisco placed the global volume of online gaming traffic at 76 petabytes per month in 2016, projected to grow to grow seven-fold by 2020 (See chart). These estimates are up sharply from values published just three years earlier.
Notably, “online gaming” is one of only four segments of “consumer internet traffic” data that Cisco disaggregates, the others being internet video, web/email/data, and file sharing, and is the fastest-growing at 47%/year. Gaming devices are also used for other cloud-based activities such as web-browsing and video streaming.
Source: Cisco Visual Networking Index
Some gaming equipment manufacturers are exploring the establishment of data centers dedicated solely to supporting gaming. As described in the Trendsetter interview at the top of this issue, NVIDIA’s Geforce NOW service offers high-end virtual gaming PCS which can be “rented” by gamers seeking a peak gaming experience.
Results are coming in from our bench testing. Among these are nine windows-based desktop PCs and five laptops carefully chosen to characterize the spectrum of gear used by gamers in the marketplace (detailed descriptions in our latest report, summary below under “Project Publications”). The development of these specs and granular results were reported in Issue 2 and Issue 3 of this newsletter.
Initial results in active mode include two test procedures, with average power during the gaming cycle varying by a factor of seven (~50W to ~425W) for the desktops and a factor of four (from ~40 to ~150W) for the laptops (see chart). 3DMark’s Firestrike benchmark is chosen as the stress test because it can run on the full range of platforms and scales and tracks well. For comparison, Skyrim TES, also desirable because of its popularity and that it is compatible with most of PC and console platforms, is run by a member of our team under more real-world conditions using a highly scripted routine for replicability.
The E-series systems are entry-level, “M” are mid-range, and “H” are high-end. The L-series are laptops, in order of increasing computing power. Values shown are average system power over the test period. Specs for each machine are recorded here. Note that these results do not include display power. Simulated benchmark is Firestrike, except for Macs (M1 and L3) with run Unigine Valley (Note: This chart updated May 26, 2018)
In some cases, the choice of test method is clearly a big factor, particularly for the machines at the more powerful end of the spectrum. Note that system H2 (the Digital Storm - Velox machine) is our highest-performing machine, yet under Skyrim uses less energy than all but the simplest entry-level machines.
We are also looking closely at non-active-gaming modes. As seen in the following chart power use there can be significant and, interestingly, follows a qualitatively different pattern across the systems than power during gaming. For example, as expected, the lower-end entry level machine E3 uses far less power than the high-end H2 machine in non-active mode than active mode, yet far more during gaming at least in the case of the selected benchmark. Idle power also varies considerably in relation to more compute-intensive modes as well as active gaming. In most cases, idle power consumption is only slightly lower than that for browsing or streaming, and in many cases draws significantly more than half the power of active gaming mode. These latter findings are presumably a reflection of power management strategies with varying levels of effectiveness.
E-series are entry-level systems, “M” are mid-range, and “H” are high-end. The L-series are laptops, in order of increasing computing power. Values shown are average system power over the test period. Specs for each machine are recorded here. Note that these results do not include display power.
Notable Industry Activities
Arising from a policy recommendation in a 2009 study, all gaming console manufacturers (Sony, Microsoft, and Nintendo) became engaged with the European Commission in developing a voluntary agreement on improving the energy efficiency of their products. The parties adopted a “self-regulatory approach”, which they see as more effective and adaptable than formal regulation.
Per the official website, “[t]he Voluntary Agreement specifies commitments industry must make regarding maximum power limits and auto-power down for different types of mains-powered game consoles 'placed on the market' within EU countries (except those consuming under 20 W). Commitments made under the Voluntary Agreement will improve game console energy efficiency without compromising console performance and the gaming experience. Gamers will also benefit by receiving additional information on the energy consumption of their consoles and instructions on how to minimise energy consumption.”
The group’s most recent meeting took place this past July, at which a wide array of new data and analyses were presented, showing that significant strides are being made in the energy use of game consoles. Note that energy intensities tend to be significantly lower than most gaming PCs. Among the findings emphasized from a report on benchmarking also presented at the meeting, “[t]he complexity of these devices makes it difficult to define computational output in a way that can be accurately, consistently, and correctly compared across game consoles and PC gaming machines. … it’s unlikely that a benchmark for active gaming will ever be good enough on which to base efficiency regulations or utility incentives to promote more efficient products.” Among the points the authors emphasize is that there are many measures of gaming performance and user experience beyond simple framerates, and that is it is difficult or impossible to achieve repeatable, representative, normalized, and comparable benchmark metrics.
Virtual reality has gained considerable interest among gamers, with several manufacturers bringing products to the market for gaming computers and consoles. Initial consideration suggests potential intrinsic energy savings, due to the far smaller active display area. However, the VR technology requires much higher framerates than two-dimensional displays, thus placing greater computing demands on the gaming system. Moreover, 2D displays are often used in conjunction with VR for orientation and to enable others in the room to follow the gaming session.
We have produced the first publicly available measured data on gaming computer and console energy use under VR. Figure xxshows active gaming power over a scripted gaming session for two VR games—Project Cars and Batman Arkam— on two popular VR technologies: Oculus and Vive, as played on our three desktop and one laptop systems. In the case of Batman Arkam, the VR system was run under full-resolution and foveated-rendering modes. The variations in energy use between 2D and VR are notable. The direction of change varies, ranging from an increase of 38% (93W, System H2 running Batman Arkam) to a reduction of about 15% (52W, System H1 running Batman Arkam with foveated-rendering mode). Averaged across all the systems power in active mode was 22% (53W) higher for Batman Arkam with full-resolution. When foveated rendering is activated, power drops by 5% (9W) compared to the 2D display. Average power for Project Cars is 5% (13W), which appears to be using Foveated Rendering by default.
Manufacturers of graphics processing units have several reasons to be interested in lightening the computing power requirements for VR, most notably that the path to cordless headsets requires less data transfer. An emerging strategy for accomplishing this is to gradually reduce the rendering fidelity along a gradient from the center of view to the periphery of vision, as the eye’s fovea is most sensitive in the central area. Some games are beginning to implement this strategy, which we have measured in the case of Batman Arkham, on our H2 system. This game refers to the two modes as “Full Resolution” and “Fixed-Foveated”, the latter referring to a rendering gradient where resolution drops off towards the periphery. We identified a remarkable 30% reduction in average power across the Fixed-Foveated gaming session for both Oculus and Vive. We conducted blinded back-to-back “A|B” tests with three gamers, only one of whom eventually noticed the reduction of fidelity at the periphery of their vision. Batman’s implementation allows users to vary the size of the high-resolution central area, as well as pixel density and pixel quality in the down-rendered outer region.
Note: corrections have been made to this article since original publication, when VR energy was observed to be higher than 2D in preliminary test results later deemed to be inaccurate.
There is ample evidence of an extremely dynamic and rapidly changing gaming market environment, where capabilities, performance, energy use, and efficiency metrics are undergoing rapid evolutionary change. We pulled data from a pair of articles from TechSpot (here and here) to draw the charts below illustrating this with information on six generations of graphics cards, each evaluated on an otherwise identical system although with varying benchmarks.
An important observation for this particular comparison is that absolute energy use has not changed much over time, while efficiency (crudely proxied as FPS/W) has steadily increased. In essence, power-reduction gains have been “reinvested” in increased performance. That said, these benchmarking results are highly dependent on which equipment is compared, other componentry in the machine (CPU, motherboard, power supply, etc.), settings chosen for the system as well as in-game options, and the benchmarking method selected. For example, in a different comparison of two boards by AMD, the system with the R9 380 was found to use 32% less energy than the R9 480. More about this in Issue #3 of this newsletter.
Games used for benchmarking and power measurements vary in this comparison. The AMD tests are noted as having been done with “medium quality” setting; no corresponding notation was provided for NVIDIA. NVIDIA performance benchmark was made using a 1366x768 display, while the display was not specified for AMD. Anti-aliasing was noted as set to off for NVIDIA case. These are two specific families of cards and the results are not necessarily indicative of the broader trends in the marketplace.
In the first major report from our project, “An Energy-focused Profile of the Video Gaming Marketplace”, pulls together a disparate data and a wide-ranging literature on market factors that influence the energy use of gaming in the past, present, and future. The report was prepared for the California Energy Commission, sponsor of LBNL’s Green Gaming research project.
Gaming has become a major social and technological phenomenon, engaged in by a third to a half of humanity. The associated energy use has been understudied, and passed over in most energy policy and planning initiatives. The report focuses on available energy-relevant information on the non-battery-powered video gaming market, including associated technology trends and gaps in the consumer information environment.
In the report, we develop a specific profile of the California marketplace for the purposes of performing energy analysis. The resulting analytical platform is based on best-available data and industry expert opinions. The profile includes an array of 25 individual gaming systems, operated by four user types across multi-step duty cycles, and running a representative assortment of game titles. This market segmentation spans the spectrum of gaming experience, system performance, and power requirements, and is leveraged to develop a characterization of the installed base of gaming equipment and its use. We find that there are currently more than 15 million video-gaming devices in use in California (the geographic focus of the study), excluding the mobile category. While the absolute number of devices is projected to decline somewhat in response to the increasing popularity of mobile gaming, the mix of platforms and their applications is shifting towards increasingly energy-intensive configurations.
The report also sheds light on significant energy efficiency improvements occurring in the marketplace and other drivers of energy demand. In subsequent phases of this project, the specified gaming systems will be bench-tested and the results used to generate aggregate baseline energy demand assessments for California and scenarios for the future.
Comings & Goings
California Energy Commission (CEC) meeting of plug loads researchers: Over the past year, the CEC has funded a portfolio of projects at multiple research institutions to explore energy efficiency opportunities in various plug loads. Our green-gaming project is one of them. In October, the research teams gathered in Sacramento to share early results. Projects included development of user interfaces to facilitate power management, development of new apps and smart context-aware power outlets, streamlining methods to enable devices to report their own energy use to the network, and on-board storage of energy harvesting from ambient sources (e.g. IR signals) to zero out standby power requirements from the grid.
GPU Technology Conference: We attended the GTC 2017 conference, which as little as two years ago was a principal venue for the latest gaming technology. While the event still features some gaming hardware innovations, it has rapidly become more focused on GPU data center technologies as well as products in the rapidly growing Artificial Intelligence and Deep Learning spaces. However, NVIDIA products supporting Virtual Reality applications computer gaming and other emerging industries (occupational training, for example) were extremely well represented. We were able to observe the latest eye tracking technology that can improve the performance of Foveated Reconstruction for VR headset displays (see article above on “Emerging Technologies” for our preliminary measurements), which holds the promise of providing large energy savings due to a reduced GPU rendering load. Our team is in contact with the principal NVIDIA scientists developing this exciting new technology and energy performance measure are underway. More about this in an upcoming issue of this newsletter.
Green-up Your Game
Tuning up your existing rig: Choose the most efficient anti-aliasing algorithm available (MFAA or similar).
When buying gear: Specify motherboards with switch-level controls (for empty slots), low power-delivery losses, power-connector quality.
More gamer tips here.
* * *
You’ll find lots of information about green gaming at our website.
Send feedback and suggestions of topics you'd like to see us cover to: Evan Mills