How To Cut Cloud Gaming Bitrates In Half So That Twice As Many Users Can Play

TL;DR: Beamr CABR operating with the Intel Media SDK hardware encoder powered by Intel GPUs is the perfect video encoding engine for cloud gaming services like Google Stadia. The Intel GPU hardware encoder reaches real-time performance with a power envelope that is 90% less than a CPU based software solution. When combined with Beamr CABR (Content-Adaptive Bitrate) technology, the required bandwidth for cloud gaming is reduced by as much as 49% while delivering higher quality 65% of the time. Using the Intel hardware encoder combined with Beamr CABR enables players to enjoy a gaming experience that is competitive to a console and able to be streamed by cloud gaming platforms. Get more information about how CABR works.

The era of cloud gaming.

With the launch of Google Stadia, we have entered a new era in the games industry called cloud gaming. Just as streaming video services opened media and entertainment content to a broader audience by freeing it from the fixed frameworks of terrestrial (over-the-air), cable, and satellite distribution, so to will cloud gaming open gameplay to a larger audience. Besides extending gameplay to virtually anywhere the user has a network-connected device, the ability for a player to access an extensive library of games without needing to use a specific piece of hardware will push 25.9 million players to cloud gaming platforms by 2023, according to the media research group Kagan.

In addition to opening up gameplay to an “anywhere/anytime” experience. A major user experience benefit of cloud gaming is that players will not necessarily need to purchase a game, but in many cases will be free to access a vast library of their choosing instantaneously. Cloud gaming services promise the quality of a console or PC experience, but without the need to own expensive hardware and the configuration and software installation work that comes with that. 

The one constraint that could cause cloud gaming to never catch up with the console experience.

With the wholesale transition of video entertainment content from traditional broadcast and physical media to streaming distribution, it is not hard to project the same pattern will occur for games. Except now, unlike the early days of video streaming where a 3Mbps home Internet connection was “high speed,” and the number of devices able to decode and reliably play back H.264 video was limited, even the lowest cost smartphone can stream video with acceptable quality. 

Yet, there is a fundamental constraint that must be overcome for cloud gaming to reach its full market potential, and that is the bandwidth required to deliver a competitive video experience at 1080p60 or 4kp60 resolution. To better understand the bandwidth squeeze that is unique to cloud gaming, let’s examine the data and signal flow. 

In FIGURE 1 we see the cloud gaming architecture moves compute-intensive operations, like the graphics rendering engine, to the cloud.

FIGURE 1

Shifting the compute-intensive function to the cloud eliminates device technical capability from being a bottleneck. However, as a result of the video rendering and encoding function not being local to the user, it means the video stream needs to be delivered over the network, with latency in the tens of milliseconds. And, at a framerate that is double the entertainment video frame rate of 24, 25, or 30 frames per second. Additionally, video game resolutions need to be HD with 4K preferable. Also, HDR is an increasingly important capability for many AAA game titles.

None of these requirements are impossible to meet, except as a result of needing fast encoding speed, the encoder must be operated in a mode that makes it difficult to produce high-quality and with small stream size. Because of the added time needed for the encoder to create B frames, and without the benefit of a look-ahead buffer, producing high quality with low bitrate is not possible. Hence why cloud gaming services require a significantly higher bitrate than what is possible with traditional video on demand streaming video services.

Beamr has been innovating in the area of performance, allowing us to encode H.264 and HEVC in software with breathtaking speed, even when running our most advanced Content-Adaptive Bitrate (CABR) rate-control. For video applications where a single encoder can serve hundreds of thousands or even millions of users, the compute requirement to do this in software, given the tremendous benefits of lower bitrate and higher quality, makes it easy to justify. But, in an application like cloud gaming, where the video encoder is matched 1:1 to every user, the computing cost to do this in software makes it uneconomical. The answer is to use a hardware encoder controlled by software, and running a content-adaptive optimization process which can deliver the additional bitrate savings needed.

FIGURE 2 illustrates the required Google Stadia bitrates.

FIGURE 2

The answer is to leverage hardware and software.

The Intel Media SDK and GPU engines occupy a well-established position in the market, with many video services relying on its included HEVC hardware encoder for real-time encoding. However, using the VBR rate-control only, there is a limit to the quality available when bitrate efficiency is essential. The advantage of Beamr’s next-generation rate-control technology, CABR (Content-Adaptive Bitrate), combined with Intel GPUs, is the secret to delivering bitrate efficiency and quality, in real-time, with 90% less power than software alone. 

In verified testing, Beamr has shown that the Intel Media SDK hardware encoder controlled by CABR will produce the same perceptual quality as VBR encodes, with a confidence level greater than 95%. Using CABR gives a meaningful impact on user experience. 65% of the time, the player will perceive better quality at the same bandwidth, even while the gaming platform experiences up to a 49% reduction in the bandwidth required to provide the same quality level.

Watch Beamr Founder Sharon Carmel present Beamr CABR integrated with Intel Gen 11 hardware encoder at Intel Experience Day October 29, 2019 in Moscow.

Proof of performance.

As an image science company, Beamr is committed to proof of performance with all claims. For this reason, the industry recognizes that all technology, products, and solutions which carry the Beamr name, represent the pinnacle of quality. For this reason, it was insufficient to integrate CABR with the Intel Media SDK without being able to prove that the original quality of the stream is always preserved and that the user experience is improved. Testing comprised corresponding 10-second segments extracted from clips created with the Intel hardware encoder using VBR, and clips encoded using the Intel hardware encoder but with the integrated Beamr CABR rate-control. 

The only way to test perceptual quality is with subjective techniques. We used a process similar to forced-choice double stimulus (FCDS), and closely approximating the ITU BT.500 method. Using the Beamr Auto-VISTA framework, we recruited anonymous viewers from Amazon Mechanical Turk where each viewer was shown corresponding segment pairs and asked to select which video had lower quality. The VBR and CABR encoded files were placed at random on the left and right sides. Validation pairs were used to verify the user’s capabilities with visible artifacts inserted, and only test results for users who correctly answered all four validation pairs were incorporated into the analysis. The viewers had up to five attempts to view the pairs before making a decision. Each viewer watched 20 segment pairs consisting of sixteen actual CABR, and VBR encodes, and four validation pairs.

Games used for testing were: CSGO, Fallout, and GTA5. To reflect realistic bitrates, we only tested the middle four bitrates out of the six bitrates provided. This was because the bitrate for the top layer was very high, and the bottom layer quality was very low. The four bitrates tested were spaced one JND (just noticeable difference) apart. Each target test pair was viewed 13 to 21 times by valid users, with a total of 800 target pair viewings, or about 17 viewings per pair on average. The total number of valid test sessions were 50, completed by more than 40 unique viewers. 

Peeling back the data, you will notice that the per-pair statistical distribution is quite symmetrical above and below 50%. With the sampling base, this phenomenon is no surprise; human perception varies. The overall results had 800 views of 48 pairs, which make the statistical certainty higher, indicating that CABR is not compromising perceptual quality. 

FIGURE 4 shows CABR encodes had the same perceptual quality as VBR and with a confidence level of more than 95%.

FIGURE 4

Better quality, lower bitrate.

Beamr CABR encoded streams offer higher quality when compared subjectively to a VBR equivalent encode, while offering a bitrate savings of up to 49%. Benefits of CABR for cloud gaming or any live streaming service, are quantified by better quality, greater bandwidth savings, and a reduction in storage cost. For the files that we tested, the aggregated metrics were as follows:

  • 65% of the time, users will experience better quality for a given bandwidth.
  • 40% bandwidth savings on average across all three titles (GTA5 had a savings of 49%).
  • 30% overall storage savings.

FIGURE 5, 6, and 7 illustrate for the three video samples used that for a given User Bandwidth, CABR provides higher quality. You will interpret the chart by observing that where VBR is blue, CABR is BLACK (higher quality), and where VBR is turquoise, CABR is BLUE.

FIGURE 5
FIGURE 6
FIGURE 7

Conclusion.

Beamr CABR controlling the Intel Media SDK hardware encoder is the perfect video encoding engine for cloud gaming services like Google Stadia. The Beamr CABR rate-control and optimization process works with all Intel codecs, including AVC, HEVC, VP9, and AV1. All bitstreams produced by the Intel + Beamr CABR solution are fully standard-compliant and work with every player in the field today. Beamr CABR is proven and protected by 46 International patents, meaning there is no other solution that can reduce bitrate by as much as 49% while working in real-time using a closed-loop perceptually aligned quality measure to guarantee the original quality. 

The single most important technical hurdle for anyone building or operating a cloud gaming service or platform is the bandwidth consumption required to deliver a player experience on par with the console. Now, with Intel + Beamr CABR, the ideal solution is here; one that can reach the performance and density needed for cloud gaming at scale, so that more players can enjoy a premium gaming experience. Streaming video upended the media and entertainment business, with the rise of Netflix, Hulu, Amazon Prime Video, Disney+, Apple TV Plus, and dozens of other tier-one streaming services. In the same way, cloud gaming will create new service platforms, gaming experiences, and business models. 

To experience the power of Beamr CABR controlling the Intel hardware encoder, send an email to info@beamr.com.

Beamr AVC & HEVC Live Encoding Performance Milestones

It has been two years since we published a comparison of the two leading HEVC software encoder SDKs; Beamr 5, and x265. In this article you will learn how Beamr’s AVC and HEVC software codec SDKs have widened the computing performance gap further over x264 and x265 for live broadcast quality streaming.

Why Performance Matters

With the performance of our AVC (Beamr 4) and HEVC (Beamr 5) software encoders improving several orders of magnitude over the 2017 results, it felt like the right time to refresh our benchmarks, this time with real-time operational data.

It’s no secret that x264 and x265 have benefited greatly, as open-source projects, from having thousands of developers working on the code. This is what makes x264 and x265 a very high bar to beat. Yet even with so many bright and talented engineers donating tens of thousands of development hours to the code base, the architectural constraints of how these encoders were built limit the performance on multicore processors as you will see in the data below.

Creative solutions have been developed which enable live encoding workflows to be built using open-source. But it’s no secret that they come with inherent flaws that include being overly compute intensive and encumbered with quality issues as a result of not being able to encode a full ABR stack, or even a single 4K profile, on a single machine.

The reason this matters is because the resolutions that video services deliver continues to increase. And as a result of exploding consumer demand for video, data consumption is consuming network bandwidth to the point that Cisco reports in 2022, 82% of Internet traffic will be video.

Cisco says in their Visual Networking Index that by 2022 SD resolution video will comprise just 23.4% of Internet video traffic, compared to the 60.2% of Internet video traffic that SD video consumed in 2017. What use to represent the middle-quality tier, 480p (SD), has now become the lowest rung of the ABR ladder for many video distributors.

1080p (HD) will makeup 56.2% of Internet video traffic by 2022, an increase from 36.1% in 2017. And if you thought the resolution expansion was going to end with HD, Cisco is claiming in 2022, 4K (UHD) will comprise 20.3% of all Internet-delivered video. 

Live video services are projected to expand 15x between 2017 and 2022, meaning within the next three years, 17.1% of all Internet video traffic will be comprised of live streams.

These trends demonstrate the industry’s need to prepare for this shift to higher resolution video and real-time delivery with software encoding solutions that can meet the requirement for live broadcast quality 4K.

Blazing Software Performance on Multicore Processors

The Beamr 5 software encoder utilizes an advanced thread management architecture. This represents a key aspect of how we can achieve such fantastic speed over x265 at the same quality level.

x265 works by creating software threads and adding them to a thread pool where each task must wait its turn. In contrast, Beamr 5 tracks all the serial dependencies involved with the video coding tasks it must perform, and it creates small micro-tasks which are efficiently distributed across all of the CPU cores in the system. This allows the Beamr codec to utilize each available core at almost 100% capacity. 

All tasks added to the Beamr codec thread pool may be executed immediately so that no hardware thread is wasted on tasks where the data is not yet available. Interestingly, under certain conditions, x265 can appear to have higher CPU utilization. But, this utilization includes software threads which are not doing any useful work. This means they are “active” but not processing data that is required for the encoding process.

Adding to Beamr encoders thread efficiency, we have implemented patented algorithms for more effective and efficient video encoding, including a fast motion estimation process and a heuristic early-termination algorithm which enables the encoder to reach a targeted quality using fewer compute resources (cycles). Furthermore, Beamr encoders utilize the latest AVX-512 SIMD instruction set for squeezing even more performance out of advanced CPUs.

The end result of the numerous optimizations found in the Beamr 4 (AVC) and Beamr 5 (HEVC) software encoders is that they are able to operate nearly twice as fast as x264 and x265 at the same quality, and with similar settings and tools utilization. 

Video streaming services can benefit from this performance advantage in many ways, such as higher density (more channels per server) which reduces operational costs. To illustrate what this performance advantage can do for you- consider that at the top-end, Beamr 5 is able to encode 4K, 10-bit video at 60 FPS in real-time using just 9 Intel Xeon Scalable cores where x265 is unable to achieve this level of performance with any number of computing cores (at least on a single machine). And, as a result of being twice as efficient, Beamr 4 and Beamr 5 can deliver higher quality at the same computing envelope as x264 and x265.


The Test Results

For our test to be as real-world as possible, we devised two methodologies. In the first, we measured the compute performance of an HEVC ABR stack operating both Beamr 5 and x265 at live speed. And for the second test, our team measured the number of simultaneous live streams at 1080p, comparing Beamr 4 with x264, and Beamr 5 with x265; and for 4K comparing Beamr 5 with x265. All tests were run on a single machine.

Live HEVC ABR Stack: Number of ABR Profiles (Channels)

This test was designed to find the maximum number of full ABR channels which can be encoded live by Beamr 5 and x265 on an AWS EC2 c5.24xlarge instance.

Each AVC channel comprises 4 layers of 8-bit 60 FPS video starting from 1080p, and the HEVC channel comprises either 4 layers of 10-bit 60 FPS video (starting from 1080p), or 5 layers of 10-bit 60 FPS video (starting from 4K).

Live HEVC ABR Stack Test – CONFIGURATION

  • Platform: 
    • AWS EC2 c5.24xlarge instance
    • Intel Xeon Scaleable Cascade Lake @ 3.6 GHz
    • 48 cores, 96 threads
  • Presets: 
    • Beamr 5: INSANELY_FAST
    • x265: ultrafast
  • Content: Netflix 10-bit 4Kp60 sample clips (DinnerScene and PierSeaside)
  • Encoded Frame Rate (all layers): 60 FPS
  • Encoded Bit Depth (all layers): 10-bit
  • Encoded Resolutions and Bitrates: 
    • 4Kp60@18000 Kbps (only in 4K ABR stack)
    • 1080p60@3750 Kbps
    • 720p60@2500 Kbps
    • 576p60@1250 Kbps
    • 360p@625 Kbps

Live HEVC ABR Stack Test – RESULTS

HEVC Live Stacks Beamr 5 vs x265

NOTES:

(1) When encoding 2 full ABR stacks with Beamr 5, 25% of the CPU is unused and available for other tasks.

(2) x265 cannot encode even a single 4K ABR stack channel at 60 FPS. The maximum FPS for the 4K layer of a single 4K ABR stack channel using x265 is 35 FPS.


Live AVC & HEVC Single-Resolution: Number of Channels (1080p & 4K)

In this test, we are trying to discover the maximum number of single-resolution 4K and HD channels that can be encoded live by Beamr 4 and Beamr 5 as compared with x264 and x265, on a c5.24xlarge instance. As with the Live ABR Channels test, the quality between the two encoders as measured by PSNR, SSIM and VMAF was always found to be equal, and in some cases better with Beamr 4 and Beamr 5 (see the “Quality Results” section below).

Live AVC Beamr 4 vs. x264 Channels Test – CONFIGURATION

  • Platform: 
    • AWS EC2 c5.24xlarge instance
    • Intel Xeon Scaleable Cascade Lake @ 3.6 GHz
    • 48 cores, 96 threads
  • Speeds / Presets: 
    • Beamr 4: speed 3
    • x264: preset medium
  • Content: Netflix 10-bit 4Kp60 sample clips (DinnerScene and PierSeaside)
  • Encoded Frame Rate (all channels): 60 FPS
  • Encoded Bit Depth (all channels): 8-bit
  • Channel Resolutions and Bitrates: 
    • 1080p60@5000 Kbps

Live AVC Beamr 4 vs. x264 Channels Test – RESULTS

Live HEVC Beamr 5 vs. x265 Channels Test – CONFIGURATION

  • Platform: 
    • AWS EC2 c5.24xlarge instance
    • Intel Xeon Scaleable Cascade Lake @ 3.6 GHz
    • 48 cores, 96 threads
  • Speeds / Presets: 
    • Beamr 5: INSANELY_FAST
    • x265: ultrafast
  • Content: Netflix 10-bit 4Kp60 sample clips (DinnerScene and PierSeaside)
  • Encoded Frame Rate (all channels): 60 FPS
  • Encoded Bit Depth (all channels): 10-bit
  • Channel Resolutions and Bitrates: 
    • 4K@18000 Kbps
    • 1080p60@3750 Kbps

Live HEVC Beamr 5 vs. x265 Channels Test – RESULTS

NOTES:

(1) x265 was unable to reach 60 FPS for a single 4K channel, achieving just 35 FPS at comparable quality.


Quality Comparisons (PSNR, SSIM, VMAF)

Beamr 5 vs. x265

NOTES:

As previously referenced, x265 was unable to reach 4Kp60 and thus PSNR, SSIM, and VMAF scores could not be calculated, hence the ‘N/A’ designation in the 3840×2160 cells.

Beamr 4 vs. x264


Source Clip Downloads

All source clips used in this comparison are publicly available at  https://media.xiph.org/video/derf.


Conclusion

Video engineers are universally focused on the video encoding pillars of computing efficiency (performance), bitrate efficiency, and quality. Even as technology has enabled each of these pillars to advance with new tool sets, it’s well known that there is still a tradeoff between each that is required.

On one hand, bitrate efficiency requires tools that sap performance, and on the other hand, to reach a performance (speed) target, tools which could positively affect quality cannot be used without harming the performance characteristics of the encoding pipeline. As a result, many video encoding practitioners have adapted to the reality of these tradeoffs and simply accept them for what they are. Now, there is a solution…

The impact of adopting Beamr 4 for AVC and Beamr 5 for HEVC transcends a TCO calculation. With Beamr’s high-performance software encoders, services can achieve bitrate efficiency and performance, all without sacrificing video quality. 

The use of Beamr 4 and Beamr 5 opens up improved UX with an increase in resolution or frame-rate which means it is now possible for everyone to stream higher quality video. As the competitive landscape for video delivery services continues to evolve, never has the need been greater for an AVC and HEVC codec implementation that can deliver the best of all three pillars: performance, bitrate efficiency, and quality. With the performance data presented above, it should be clear that Beamr 4 and Beamr 5 continue to be the codec implementations to beat.