TR Forums

Video playback on computers has fascinated me since the early '90s. I can still vividly remember going to a computer show back in 1993 and seeing, for the first time, a computer playing a full screen, feature length movie. It was a 66MHz 486 playing Jurassic Park off of a CD-ROM, with the help of a massive MPEG-1 decoder sitting in one of the ISA slots. The whole spectacle made me very happy.

Today, most of the video that you'll see on a computer screen is delivered via the Internet. The resolutions, frame rates, and colour depth are all much higher now, thanks to VP9 and HEVC compression. These codecs are far more efficient than MPEG-1, making the decoding process a compute-intensive task. Phones, tablets, and newer PCs have dedicated VP9 and HEVC decoding logic built right in. However, slightly older PCs have to rely on a relatively inefficient general purpose processor to perform all of the heavy lifting.

So, how capable are CPUs at decoding VP9 video and displaying them through a kludgy web browser? I tested 19 different CPUs to find out.

First, a few notes on testing:
-I played 480p - 2160p/60 YouTube video using the HTML5 player in Chrome 60, making sure that VP9 was being streamed each time. Chrome 64 was used on the Ryzen system.
-I only used 30 and 60 fps videos. 24 fps is far too easy.
-The AMD Ryzen, A10 and Intel Atom X7 ran Windows 10; all of the other machines were tested with Windows 7 SP1.
-I would've loved to include an Athlon XP or even an overclocked Pentium III, but all modern browsers require SSE2.
-OC indicates that the processor was overclocked.

Also, I noticed that the initial "surge" of data that occurs when a video starts playing burns quite a few CPU cycles. This stresses many of the older CPUs, causing them to drop frames. I've measured downstream activity as high as 70-80Mb/s during the first 30 seconds of playback. So, to keep things consistent, I waited 30 seconds before recording any performance numbers.

With that out of the way, here are the results:



Some interesting takeaways:
-The A10-8700p is curiously slow for a 4T processor, being only slightly faster than the "little core" Atom X7 Z8700. The ancient C2D E8600 is in a completely different dimension, easily outclassing both processors despite only having 2 threads to work with.
-Speaking of the C2D, I find it amazing how an old dual-core manages to handle 1080p/60 and 1440p VP9 streams without ANY assistance from a hardware decoder.
-1080p @ 60 fps actually seems to be a little more demanding than 1440p @ 30.
-Something in the Core ix processors make them brutally fast at video decoding. The quad-core i5-3470, clocked at a boring 3.4GHz, not only demolishes the 4GHz Q6700, but also wipes the floor with the 4.07GHz Phenom X6!
-The overclocked P4-HT on LGA775 puts the old s478 P4s to shame. It's not even close.
-Just look at how much faster Ryzen7 @ 4GHz is than Phenom X6 @ 4GHz. I'm guessing that an FX-6300 would perform similarly to the Phenom. This is like going from the K6 to Athlon, all over again!

setaG_lliB wrote:

-Something in the Core ix processors make them brutally fast at video decoding. The quad-core i5-3470, clocked at a boring 3.4GHz, not only demolishes the 4GHz Q6700, but also wipes the floor with the 4.07GHz Phenom X6!

That's not terribly surprising. Intel has worked decently hard to include some decent decode hardware in more recent versions of their IGPs. Cool you got all that data. Now we just need a decent compression scenario where it doesn't take 5-10GB of your data cap to stream 2160p video for an hour.

DancinJack wrote:

setaG_lliB wrote:

-Something in the Core ix processors make them brutally fast at video decoding. The quad-core i5-3470, clocked at a boring 3.4GHz, not only demolishes the 4GHz Q6700, but also wipes the floor with the 4.07GHz Phenom X6!

That's not terribly surprising. Intel has worked decently hard to include some decent decode hardware in more recent versions of their IGPs.

The i5's IGP was disabled. It was using a GTX-680, which doesn't decode VP9 in hardware.
Actually, none of these processors had the luxury of VP9 hardware acceleration.

setaG_lliB wrote:

-The A10-8700p is curiously slow for a 4T processor, being only slightly faster than the "little core" Atom X7 Z8700. The ancient C2D E8600 is in a completely different dimension, easily outclassing both processors despite only having 2 threads to work with.

The single-channel RAM is likely holding it back. Pushing lots of pixels around requires lots of RAM bandwidth, and that's going to be in short supply on a single-channel system.

setaG_lliB wrote:

-Something in the Core ix processors make them brutally fast at video decoding. The quad-core i5-3470, clocked at a boring 3.4GHz, not only demolishes the 4GHz Q6700, but also wipes the floor with the 4.07GHz Phenom X6!

Probably due to Intel Quick Sync, which puts dedicated video decode/encode hardware on the CPU AVX support, which was introduced with Sandy Bridge.

Love the chart. Thanks for sharing.

-1080p @ 60 fps actually seems to be a little more demanding than 1440p @ 30.

It is.
2560x1440@ 30 fps is ~110 MPixels per second
1920x1080@ 60 fps is ~124 MPixels per second
It's exactly 12.5% more pixels to push, so a CPU that is running at 90% doing the former is going to be trying to push past 100% on the latter, which is exactly what you see with the A10.

Interesting stuff. I'd still definitely recommend a hardware decoder for high-resolution and high-framerate video, but it's good to know that modern CPUs can handle VP9 in software if necessary.

just brew it! wrote:

setaG_lliB wrote:

-The A10-8700p is curiously slow for a 4T processor, being only slightly faster than the "little core" Atom X7 Z8700. The ancient C2D E8600 is in a completely different dimension, easily outclassing both processors despite only having 2 threads to work with.

The single-channel RAM is likely holding it back. Pushing lots of pixels around requires lots of RAM bandwidth, and that's going to be in short supply on a single-channel system.

Plus, that memory controller SUX. Prior to Ryzen, take theoretical memory bandwidth for AMD procs and cut it in half, to start. DDR3-1866 is not PC3-14900 in the FX/Trinity lineage.

Ryzen, on the other hand, posts incredible synthetic memory benchmark numbers.

I have found the Firefox Quantum has lower CPU usage than Chrome when playing back VP9. You might want to try it on some of those low end CPUs to see if there's any difference.

I am using a Pentium EE 965 @ 4 GHz and 1080p VP9 Youtube videos only use around 20% CPU usage. The closest chip in your list is the 3.2GHz Pentium D.

Thanks for posting.

just brew it! wrote:

setaG_lliB wrote:

-The A10-8700p is curiously slow for a 4T processor, being only slightly faster than the "little core" Atom X7 Z8700. The ancient C2D E8600 is in a completely different dimension, easily outclassing both processors despite only having 2 threads to work with.

The single-channel RAM is likely holding it back. Pushing lots of pixels around requires lots of RAM bandwidth, and that's going to be in short supply on a single-channel system.

setaG_lliB wrote:

-Something in the Core ix processors make them brutally fast at video decoding. The quad-core i5-3470, clocked at a boring 3.4GHz, not only demolishes the 4GHz Q6700, but also wipes the floor with the 4.07GHz Phenom X6!

Probably due to Intel Quick Sync, which puts dedicated video decode/encode hardware on the CPU AVX support, which was introduced with Sandy Bridge.

I'm going to have to try and find a second 1600MHz SO-DIMM and re-test the A10 machine.
You might be right about AVX support. Although to be honest, I'd be pretty impressed if Chrome is using it. Didn't it take them (and other browsers) forever to finally use SSE?

Takeshi7 wrote:

I have found the Firefox Quantum has lower CPU usage than Chrome when playing back VP9. You might want to try it on some of those low end CPUs to see if there's any difference.

I am using a Pentium EE 965 @ 4 GHz and 1080p VP9 YouTube videos only use around 20% CPU usage. The closest chip in your list is the 3.2GHz Pentium D.

I retested the C2D E8600 machine with Firefox Quantum. CPU usage was indeed significantly lower, at around 4% for 1080p (Chrome was using 26-35%). However, I noticed that YouTube was streaming AVC (H.264) video to Firefox. H.264 is decoded in hardware by most video cards.

setaG_lliB wrote:

I retested the C2D E8600 machine with Firefox Quantum. CPU usage was indeed significantly lower, at around 4% for 1080p (Chrome was using 26-35%). However, I noticed that YouTube was streaming AVC (H.264) video to Firefox. H.264 is decoded in hardware by most video cards.

Interesting. On my PC both Firefox Quantum and Chrome use VP9 for Youtube. I have to use Edge or a Chrome plugin in order for h.264 to be used.

I found a way to enable VP9 on Firefox. In general, Mozilla's VP9 implementation is faster than Google's, but it appears to be using some form of hardware acceleration. Obviously not for decoding, but perhaps for rendering the decoded frames? The Pentium D and Core 2 Duo--both equipped with GeForce video cards--were much faster using Firefox. However, the third machine I tested, a Core Duo T2500 with a mobile Radeon X1600, was a complete slideshow.

VP9 @ 1080p - Chrome vs Firefox

Pentium D 935 (3.2GHz)
Chrome: 87-100% with occasional stuttering.
Firefox: 53-74%, completely smooth

Core Duo "Yonah" T2500 (2GHz)
Chrome: 84-100 with occasional stuttering.
Firefox: 18-68 but only 0.2 fps

Core 2 Duo E8600 (3.33GHz)
Chrome: 26-35
Firefox: 16-21
Chrome @ 1440p: 55-67
Firefox @ 1440p: 35-41

Dammit. My curiosity is eventually going to force me to rebench all of these machines.

setaG_lliB wrote:

Didn't it take them (and other browsers) forever to finally use SSE?

Something to do with x86-64 web browser builds? I know x86-64 has SSE/SSE2 as a requirement in place of x87.

The original AMD64 architecture adopted Intel's SSE and SSE2 as core instructions. These instruction sets provide a vector supplement to the scalar x87 FPU, for the single-precision and double-precision data types. SSE2 also offers integer vector operations, for data types ranging from 8bit to 64bit precision.
...
The proliferation of 64-bit processors has made these vector capabilities ubiquitous in home computers, allowing the improvement of the standards of 32-bit applications. The 32-bit edition of Windows 8, for example, requires the presence of SSE2 instructions. SSE3 instructions and later Streaming SIMD Extensions instruction sets are not standard features of the architecture.

https://en.wikipedia.org/wiki/X86-64

setaG_lliB wrote:

VP9 @ 1080p - Chrome vs Firefox

Pentium D 935 (3.2GHz)
Chrome: 87-100% with occasional stuttering.
Firefox: 53-74%, completely smooth

That sounds about right to me. The difference between what I see on my Pentium EE and your Pentium D is probably down to Hyperthreading and higher clock speed.

It also has a GTX 480, so I wasn't expecting your terrible result with the ATI X1600. But I don't really see high GPU usage on my GTX 480 when I'm watching VP9 videos, so I'm not sure how much load it's really taking off the CPU.

biffzinker wrote:

SSE3 instructions and later Streaming SIMD Extensions instruction sets are not standard features of the architecture.
https://en.wikipedia.org/wiki/X86-64

Perhaps true, but seems pointless to mention considering afaik you can't find a proc that supports teh 64 bitzes without SSE3.

LostCat wrote:

biffzinker wrote:

SSE3 instructions and later Streaming SIMD Extensions instruction sets are not standard features of the architecture.
https://en.wikipedia.org/wiki/X86-64

Perhaps true, but seems pointless to mention considering afaik you can't find a proc that supports teh 64 bitzes without SSE3.

Early (pre-Venice core) Athlon-64 CPUs did not support it. But those are pretty long in the tooth these days.

I doubt SSE3 or later matter much. It's my understanding that SS3 and later were designed to cater to certain niche mathematical needs, that SSE and SSE2 cover everything you'd generally want an x87 FPU to do.

Ryu Connor wrote:

I doubt SSE3 or later matter much. It's my understanding that SS3 and later were designed to cater to certain niche mathematical needs, that SSE and SSE2 cover everything you'd generally want an x87 FPU to do.

The SSE3 enhancements primarily benefit DSP-type operations (which is potentially useful for video decoding), by allowing multiple values packed into a single register to be processed simultaneously. So while it doesn't add new capabilities per se, it does allow certain types of repetitive operations to be performed faster (by doing more than one of them at a time).

setaG_lliB wrote:

I found a way to enable VP9 on Firefox. In general, Mozilla's VP9 implementation is faster than Google's, but it appears to be using some form of hardware acceleration. Obviously not for decoding, but perhaps for rendering the decoded frames? The Pentium D and Core 2 Duo--both equipped with GeForce video cards--were much faster using Firefox. However, the third machine I tested, a Core Duo T2500 with a mobile Radeon X1600, was a complete slideshow.

VP9 @ 1080p - Chrome vs Firefox

Pentium D 935 (3.2GHz)
Chrome: 87-100% with occasional stuttering.
Firefox: 53-74%, completely smooth

Core Duo "Yonah" T2500 (2GHz)
Chrome: 84-100 with occasional stuttering.
Firefox: 18-68 but only 0.2 fps

Core 2 Duo E8600 (3.33GHz)
Chrome: 26-35
Firefox: 16-21
Chrome @ 1440p: 55-67
Firefox @ 1440p: 35-41

Dammit. My curiosity is eventually going to force me to rebench all of these machines.

you're going to have to do a few now, say VP9, H.26X, with a few different browsers, and gpu's could do a software and hardware, you opened a big can of worms

setaG_lliB wrote:

Core Duo "Yonah" T2500 (2GHz)
Chrome: 84-100 with occasional stuttering.
Firefox: 18-68 but only 0.2 fps

This confuses me...the FF one has to be unwatchable, right? I mean .2 FPS is...nothing. The descriptors you use for the different browsers are strange.

DancinJack wrote:

setaG_lliB wrote:

Core Duo "Yonah" T2500 (2GHz)
Chrome: 84-100 with occasional stuttering.
Firefox: 18-68 but only 0.2 fps

This confuses me...the FF one has to be unwatchable, right? I mean .2 FPS is...nothing. The descriptors you use for the different browsers are strange.

Yes, it is unwatchable. Scrolling is also a little choppy on Firefox. My best guess is that FF is using a method of screen compositing not supported by the X1600.

Starting to feel like I forgot more about AMD procs than I remember Guess it was over a decade ago.

LostCat wrote:

Starting to feel like I forgot more about AMD procs than I remember Guess it was over a decade ago.

Aside from the one APU, the chart doesn't even include anything from AMD made in the past 7 years.

just brew it! wrote:

LostCat wrote:

Starting to feel like I forgot more about AMD procs than I remember Guess it was over a decade ago.

Aside from the one APU, the chart doesn't even include anything from AMD made in the past 7 years.

That was kinda what I was curious about, to see how AMD stacked up even with say Bulldozer and of course Ryzen.

I'm considering if AMD has a leg up on Intel for video playback on some security camera systems. Then again, the camera ecosystems in question are using H.264 and soon 265.

This explains why my wife's old XPS 14z is having such a hard time lately. With a bajillion of tabs all auto-playing video her CPU (and RAM) makes the poor laptop cry. I guess it's time to upgrade.

DragonDaddyBear wrote:

This explains why my wife's old XPS 14z is having such a hard time lately. With a bajillion of tabs all auto-playing video her CPU (and RAM) makes the poor laptop cry. I guess it's time to upgrade.

Turn off autoplay?

DragonDaddyBear wrote:

This explains why my wife's old XPS 14z is having such a hard time lately. With a bajillion of tabs all auto-playing video her CPU (and RAM) makes the poor laptop cry. I guess it's time to upgrade.

That, or you could download and install h264ify in your browser... That'll at least provide streams her GPU can accelerate.

I didn't realize there was an extension for forcing a video type. I'll look into it.

She uses Chrome and I think that disabling auto play of all media a relatively new-ish feature. I'll need to dig for the setting.

All of that said, her CPU is still apparently ancient and is due to be replaced. I think it's an i5-2430M or something like that.

Are you sure the IGP is disabled ? in some intel cpus , the video decode block is separate from the main graphics , also even if vp9 is isn't directly supported, if vp8 is , then it may be partially hardware accelerated..