Are accelerators the cure to video's power problem or just an excuse to peddle GPUs?
- Reference: 1681386971
- News link: https://www.theregister.co.uk/2023/04/13/accelerators_streaming/
- Source link:
Video is also an incredibly resource-intensive media. As reported by the Financial Times, an ammunition plant in Norway [1]found that out the hard way. The plant, which was producing munitions for the war in Ukraine, had planned to expand only to discover that there wasn't enough power. Why? Because a new TikTok datacenter was sucking up every spare watt it could find.
And while TikTok may be on uncertain ground as it faces the [2]prospect of an outright ban in the US, it's not even the largest contributor of video. That title goes to Google (Youtube) and Netflix, according to the latest Sandvine report.
[3]
So what can be done? Well, if you ask Nvidia CEO Jensen Huang, the [4]answer is accelerated computing, preferably using its GPUs to encode and decode the video on the fly, and its data processing units (DPUs) to accelerate the movement of the data as it hurtles across the intertubes to your phone or TV.
[5]
[6]
"User generated video is driving significant growth and consuming massive amounts of power," Huang said during the company's GTC event in March. "We should accelerate all video processing and reclaim that power."
Accelerated transcoding gear
The idea of using GPUs or other dedicated accelerators to transcode video is hardly new. Nvidia's diminutive P4 and T4 GPUs have been a popular choice for video streaming applications for years. Last month Nvidia [7]unveiled the L4.
The company claims an eight-L4 node can transcode more than 1,000, 720p streams at 30fps when using the P1 preset. If these numbers sound a little optimistic, that's because Nvidia is goosing the numbers a bit by using the lowest quality preset possible.
But Nvidia isn't the only company that sees an opportunity to address growing demand for streaming video and to peddle their wares in the process. Earlier this month AMD's Xilinx division [8]unveiled its new Alveo media accelerator card, the MA35D.
[9]
AMD says the card can transcode up to 32 1080p AV1 streams at 60fps while consuming just 35W under load.
Even Intel, a company that only recently entered the GPU market, has skin in the game and actually beat both AMD and Nvidia to market with an AV1-capable chip. Intel's Flex-series GPUs – codenamed Arctic Sound – were [10]announced during Hot Chips last August. The cards were designed with video in mind.
All of these parts, regardless of whether they're from Nvidia, AMD, or Intel, feature single-slot designs and TDPs under 75W (for the most part anyway), allowing for eight or more of them to be packed into a single system. And that's not the only thing that these cards have in common. All of them are marketed toward live video, cloud gaming, and AI image processing – think Twitch or live sports broadcasting.
[11]It's time to stop fearing CPU power management
[12]Why we think Intel may be gearing up to push its GPU Max chips into China
[13]When will regulators get serious on datacenter emissions reporting?
[14]Google boffins pull back more of the curtain hiding TPU v4 secrets
No silver bullet for the silver screen
Compared to conventional CPU-based software rendering, these cards might seem like the future, but it's actually not that simple. That's because most streaming services, like Netflix, don't have to deal with large volumes of low-latency video. Instead, they can spend hours converting a video master into multiple copies for each format, resolution, and bitrate they plan to serve the video up at.
The streaming service can then distribute the videos out to content delivery networks (CDNs) and serve them up from there. And since software transcoding, while slower, tends to render higher quality video for a given bitrate, and storage is relatively cheap compared to a bunch of GPU nodes, the economics are likely to continue to favor this approach for conventional video.
[15]
Software encoders will remain relevant despite the gains in accelerated computing, Sean Gardner, head of video strategy at AMD, told The Register .
"Software encoders continue to evolve ... For some applications – some use cases – [these cards] will be good enough to finally be considered for some of those file-based, not-real-time [applications]. But for Netflix? I don't think so."
AV1 to the rescue
Despite what Jensen would have you believe – remember he wants to sell you GPUs – accelerated computing isn't the be-all-end-all of video.
With that said, accelerated computing isn't the only way to improve the efficiency of video platforms. The bandwidth consumed getting the video to your phone or TV is another major factor.
"Typically it's somewhere between 6 and 9 percent of their revenue that they spend on bandwidth," he said of AMD's streaming video customers. "One of the biggest power consumers is actually the communication side, the network that that delivers those, those Ethernet packets to the user."
This is where the Alliance for Open Media's AV1 codec comes in. AV1 is a relatively new codec that has a lot of people excited, in part because it's royalty free, but also because it's exceptionally space efficient.
Various tests have [16]shown AV1 to be anywhere from 20 to 40 percent more efficient than compared to popular web streaming codecs, including H.265.
For a platform like YouTube or Netflix, the bandwidth savings would be considerable, which is probably one of the reasons why nearly every major streaming service is a member of the Alliance for Open Media.
However, despite having been around since 2018, AV1 remains in its infancy. Only the latest generation of GPUs from AMD, Nvidia or Intel support full AV1 encode/decode. Meanwhile, Apple has yet to add support for the codec on its devices. But given AV1's robust industry backing, this particular problem is really more of a growing pain than anything else. ®
Get our [17]Tech Resources
[1] https://www.ft.com/content/f85aa254-d453-4542-a50e-fa1171971ab0
[2] https://www.theregister.com/2022/12/14/tiktok_ban_proposed_us/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/energyefficientdatacenters&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZDgnJQt8x102pkD9WS8dMwAAAEE&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://www.theregister.com/2023/03/21/nvidia_sutainability_accelerator/
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/energyefficientdatacenters&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZDgnJQt8x102pkD9WS8dMwAAAEE&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/energyefficientdatacenters&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZDgnJQt8x102pkD9WS8dMwAAAEE&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2023/03/21/nvidia_h100_l4/
[8] https://www.amd.com/en/newsroom/press-releases/2023-4-6--amd-launches-first-5nm-asic-based-media-accelerat.html
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/energyefficientdatacenters&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZDgnJQt8x102pkD9WS8dMwAAAEE&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://www.theregister.com/2022/08/24/intel_flexes_arctic_sound_gpus/
[11] https://www.theregister.com/2023/04/12/cpu_power_management_feature/
[12] https://www.theregister.com/2023/04/11/intel_gpu_china/
[13] https://www.theregister.com/2023/04/11/datacenter_emissions_regulation/
[14] https://www.theregister.com/2023/04/06/google_tpuv4_hardware_nvidia/
[15] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/energyefficientdatacenters&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZDgnJQt8x102pkD9WS8dMwAAAEE&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[16] https://streaminglearningcenter.com/codecs/netflix-on-av1.html
[17] https://whitepapers.theregister.com/
Re: This makes no sense
“bigger roads just attract more cars”
[eyeroll]
The argument the anti-car fanatics actually make is that bigger roads worsen congestion.
Also if you can save bandwidth and either:
A: deliver increased video quality for the same bandwidth utilisation, or
B: deliver the same video using less bandwidth
Then surely that's a good thing - whether or not usage will increase in the long term. And if usage increases, that demonstrates that things are good ( just like how increased train use after privatisation compared to falling train use and the Beeching cuts before it proves that privatisation was successful ).
The question surely is whether the fancy card costs more than it saves (B), or whether the increased video quality is worth the price of the fancy card (A).
Re: This makes no sense
“bigger roads just attract more cars”
well, duh!! "bigger" (actually wider) roads = more throughput = less traffic, making it more attractive to use private car vs public transport, so all other things being equal, it will obviously attract more cars. Whether that will, in fact make overall congestion worse depends on many many factors particularly whether you are widening just one road without widening feeder roads.
Anyway, to return to the bandwidth....
Bandwidth required for 4k streaming is about 25Mbps using already-current coding/decoding. Maybe you could for example code a video stream at source even better... but the current algorithms are already pretty good, I doubt you could squeeze more than 5%, maybe 10% bandwidth reduction. Then you have to encode it once at the server end and decode it 100 times at the user end, so each user device has to support both the new hardware requirements and software encoding. There's not really much rationale to need a new step-change-more-powerful GPU in every user device to save what is, ultimately, not a lot of bandwidth.
Just allow standards to gradually improve organically as they mostly always have.
Re: This makes no sense
I wasn’t suggesting that aiming for better video compression wasn’t a good thing. I just don’t see it as coming anywhere close to solving the issues being discussed
And forget about the car thing - it’s not important
Transcode?
A server transcoding is a waste of energy when it can house all resolutions. I'm sure billion dollar companies can or already do this.
The client is where the problem is. VVC/h266 certainly can help but, not when nearly no clients currently support it (besides some few OTA options).
Disk space is cheap, even if Netflix needs to have 50+ copies of the same movie to accommodate every bitrate, resolution and codec they stream it to end users in, its a pretty trivial cost in storage space and means that the end users should get the best quality video if its been encoded correctly.
Where as from my experience GPU encoded videos tend to be larger in file size and yet less quality than the same source material encoded using a CPU. So there seems little benefit to anyone other than those streaming a live event to use a video accelerator to encode on the fly.
This makes no sense
Buying very expensive GPUs to transcode your video (the article doesn’t really say transcoding from what to what) might save you some network bandwidth and storage but even using AV1, that is only going to gain you a year or two before the quantity of traffic bumps up to fill the space - a bit like the argument of “bigger roads just attract more cars”.
And the servers are still going to guzzle mega watts of power.
So yes - this sounds like a sales pitch to flog GPUs and not an even vaguely useful “solution” to the problems cited