CPU for video encoding

[BallaTheFeared]

Spot on with this comment.

QuickSync should really be used for 480P or less source material on large screens, and anything goes for pads/tablets/phones.

Haswell is debuting HQ quicksync, but that won't be around until 2013. NOT trying to derail the thread, just bringing this up. Whether its competitive with x264 remains to be seen...

You wouldn't wait a few more months, this is October you know right?

 

[exar333]

You wouldn't wait a few more months, this is October you know right?

Not sure what you are asking. I am not waiting for anything...???

 

[Johnnie.it]

Food for thought:
For encoding, BD is competitive, but I would definitely not recommend an 8150 over an i5 for gaming.

Look, I already said that with a 6950 this pc is already handling anything I've ever thrown at it.
Dirt 3 (btw, dirt showdown is bundled with the 8150), deus ex human revolution, shogun 2 total war, fear 3, the ludicrous frame-eating machines that are the space complex in x3, all at 1920x1080, max details.
I'm not upgrading for gaming. That would really be pointless. Right now, the bottleneck gaming-wise is just the screen I'm using. Unless I get my hands on a gazillionp monitor, I won't ever need anything more for gaming.

I'm upgrading for encoding movies. Which reminds me, has anybody got any experience with avivo (ati gpu-assisted transcoding)?
I've done some research, found some info, but not much practical.
All in all, it looks like gpu-assisted transcoding is not really viable right now, either with intel, nvidia or amd/ati.

In case, which forum need I pester to get some info on that?

...sorry for the OT

 

[Makaveli]

Not sure what you are asking. I am not waiting for anything...???

I think he means wait for haswell.

 

[Idontcare]

I'm upgrading for encoding movies. Which reminds me, has anybody got any experience with avivo (ati gpu-assisted transcoding)?
I've done some research, found some info, but not much practical.
All in all, it looks like gpu-assisted transcoding is not really viable right now, either with intel, nvidia or amd/ati.

In case, which forum need I pester to get some info on that?

...sorry for the OT

You want to ask the Video Cards and Graphics forum that. But speaking from personal experience I will just tell you that I use TMPGEnc which has CUDA-enabled filters (I transcode a ton of my kids DVD's and like to filter the graininess out of them - temporal noise filter FTW) and it definitely works well with an Nvidia card.

Transcoding itself sucks with GPU-enabled solutions, your image quality will suffer, but the filters used along with transcoding can be gpu-enabled without IQ loss but with great speed-up results.

 

[LoveMachine]

Look, I already said that with a 6950 this pc is already handling anything I've ever thrown at it.
Dirt 3 (btw, dirt showdown is bundled with the 8150), deus ex human revolution, shogun 2 total war, fear 3, the ludicrous frame-eating machines that are the space complex in x3, all at 1920x1080, max details.
I'm not upgrading for gaming. That would really be pointless. Right now, the bottleneck gaming-wise is just the screen I'm using. Unless I get my hands on a gazillionp monitor, I won't ever need anything more for gaming.

I'm upgrading for encoding movies. Which reminds me, has anybody got any experience with avivo (ati gpu-assisted transcoding)?
I've done some research, found some info, but not much practical.
All in all, it looks like gpu-assisted transcoding is not really viable right now, either with intel, nvidia or amd/ati.

In case, which forum need I pester to get some info on that?

...sorry for the OT

Johnnie, Anandtech did an article on this a few months ago, and I'm too lazy to go find it, but longstory short is that ANY hardware accelerated transcoding resulted in suboptimal quality. Software is still king if you are wanting high quality 1080p near-bluray level transcodes.

Also, my story mirrors yours to some extent. I don't game at all, and transcoding with Handbrake is really the only hardware intensive thing I do. So all the AMD bashing some here are posting really doesn't apply. If you have a board that can handle bulldozer, get it. This seems very logical to me, since you are getting slightly better performance vs. a 3570K, at a good cost savings. My old system needed a total upgrade, so in my case I went 3570K since I had the means and I'm a nutcase for efficiency. In your situation, BD is a completely valid option, and if I were in your shoes I'd probably get an 8120, or perhaps wait just a bit longer and get a newer Piledriver.

 

[Johnnie.it]

Transcoding itself sucks with GPU-enabled solutions, your image quality will suffer, but the filters used along with transcoding can be gpu-enabled without IQ loss but with great speed-up results.

@LoveMachine I read that article, and that's why I was talking about "GPU *assisted* transcoding". I was actually referring to what Idontcare was saying.
That encoding run that took more than 15 hours did so because I turned on every possible filter...yeah, I know, it was a dumb thing to do, but I just wanted to test that, and besides...I got a couple of 10gb files that looked EXACTLY like the bluray!!!

Anyway, I was thinking more like letting the cpu doing the actual transcoding work, and the gpu the "grunt" work. Or any combination of that, really.
You know...just let them help each out other like two old drunk buddies when they're getting thrown out of a pub.

 

[lopri]

If your processor will support them, a Phenom II X6 would be the best bet. Amazon still has the 1045T in stock, and if you shop around you should be able to find a 1090T in stock somewhere.

Seconded, under the assumption the OP is willing to invest a couple of days on overclocking and tweaking the CPU-Memory confguration.

The mainboard I'm using is an ASUS M4A89GTD PRO/USB3. I'm wondering, in case I choose an FX-8150, if I should replace it.
On the CPU support page, this CPU is reported as "Beta support only"
Ref. http://www.asus.com/Motherboards/AMD_AM3/M4A89GTD_PROUSB3/#CPUS
Well, in case it's not a problem to waste 60-70€ on a new mainboard. I was considering the ASROCK 970 PRO3.

Anyway on the benchmark section here on anandtech the two processors are not that different. One is better on first pass, but the 8150 rocks on the second pass. I didn't average the FPS but, as long as the price difference goes, I thought the 980 was better value.

Anyway, the performance right now is quite right (of course, I'm talking about everything else but encoding) so the upgrade is just for that.
Needless to say, I don't want, in any way, to save time for video encoding while sacrificing any FPS for gaming.

Also, I might consider moving to a combo intel cpu+mainboard, as long as I don't have to sell a kidney

If you don't have anything else to say about that, I'll go with ASROCK 970 PRO3 (or possibly an ASROCK 970 EXTREME3 - just 10€ price difference, but as long as I have 1 or 2 sata3 and usb3 ports, I'm fine) and an FX-8150.

Any last thoughts?

@inf64: why the 8120? Why not a 6200, if it's just to save money?

There is nothing stopping you from changing the platform to newer Intel chips, but there is nothing wrong with saving time and money by drop-in CPU upgrade, either.

I can't speak for Bulldozer performance, but I can for the motherboard you have now, and the one you have in mind. Make it short: No. Your ASUS 890GX board is a better quality board than the ASRock 970 board. The only real difference between 8 series boards and 9 series boards are Bulldozer support. Everything else basically is the same. If you're staying with Phenoms, stay with the ASUS board.

If you are going to upgrade to Bulldozer, then you may want to consider 9 Series boards, but I would still avoid that ASRock board. It is a very cheaply made board. "Extreme," "X" nomenclature may entice your eyes for a second, but that's basically all there's to it. If you are not into overclocking, Gigabyte 970X boards would be a good choice. ($80~$100) If you are going to overclock, at least step up to ASUS 990X territory. ($100~$120)

ASRock boards are great jack-of-all-traders, but those low-priced ones indeed are quite fragile. I would not subject them under stressful long-term usage.

Said that, the board you have is an excellent overclocker, especially for X6's. I have been running an 1090T and 1045T @4.0 GHz, up to 3.0 GHz CPU NB, and up to 300 HTT, for almost 2 years now. With all power saving features enabled. Only issue I have is USB 3.0 losing its connection occasionally but that has been the problem since I got it, when USB 3.0 was just being adopted.

So again, that ASRock board is a no-no. I would say your options are:

1) Keep the 890GX, grab a new 1045T or a used 1090T for less than $100 and overclock it -> Cheapest and easiest
2) Upgrade to Sandy/Ivy bridge platform. -> Costlier in both time/money spent but you get lower power consumption with better single-thread performance
3) Upgrade to Bulldozer/9 series platform -> Questionable/inconsistent performance depending on what you do. If you choose to go that route, I would still discourage you from picking that ASRock 970 board.

I have been using 1090T/1045T @4.0 GHz (3.0 GHz CPU-NB) and 2500K @4.8 GHz side-by-side for quite some time. Obviously 2500K is the faster one when it comes to single-threaded performance, but for games (which are where single-threaded performance matters) X6 can keep up with it by CPU-NB/L3 overclock. And there is a certain advantage that 6 cores can provide over 4 cores (HT or not), if you're a multitasking-type.

If you're talking about purely multi-threaded encoding performance (not multi-tasking), my experiences are:

1) DVDFab -> 2500K wins
2) Handbrake -> 1090T/1045T win
3) x264 bench -> Tie (faster 1st pass for 2500K, faster 2nd pass for 1090T/1045T)

Note that if you spend extra $100 for HT (2600K/3770K), 1090T/1045T will lose in most multi-threaded apps to 2600K/3770K, leaving multitasking as their sole (but not insignificant) advantage.

 

[Idontcare]

@LoveMachine I read that article, and that's why I was talking about "GPU *assisted* transcoding". I was actually referring to what Idontcare was saying.
That encoding run that took more than 15 hours did so because I turned on every possible filter...yeah, I know, it was a dumb thing to do, but I just wanted to test that, and besides...I got a couple of 10gb files that looked EXACTLY like the bluray!!!

Anyway, I was thinking more like letting the cpu doing the actual transcoding work, and the gpu the "grunt" work. Or any combination of that, really.
You know...just let them help each out other like two old drunk buddies when they're getting thrown out of a pub.

I see I inadvertently linked you to the now deprecated TMPGEnc version 4. Version 5 is the one you want to look into.

I also see that I need to upgrade from Ver4 to Ver5 as Ver5 now claims to support actual H.264 encoding (not just the filters) with Cuda.

NVIDIA CUDA H.264 Encoding.
Not Just for Filtering Anymore.
In addition to CUDA-enabled filter processing and decoding, TMPGEnc Video Mastering Works 5 supports H.264/AVC file output using the CUDA technology provided by NVIDIA. The CUDA technology makes use of the GPU (graphic board CPU) with multi-core parallel processing to process complex calculation problems in a short time. By splitting the processes into parallel elements, and using the massive processing power, the program can run even more effectively. Using this function with a compatible graphic board improves processing acceleration when compared with CPU-only use in most cases.

Interestingly enough they also claim to have tapped into the quicksync units on SB and IB to enable H.264 encoding as well.

Intel® Media SDK H.264 Encoding and Decoding.
Sandy Bridge and Ivy Bridge Support.
TMPGEnc Video Mastering Works 5 also uses the Intel® Media SDK (2.0) to encode and decode H.264 content. This function, which will use the new Intel Sandy Bridge micro-architecture embeded in CPUs that are planned for release in the begining of 2011, enables hardware support of the encoding and decoding process. Tests conducted with TMPGEnc Video Mastering Works 5 showed an acceleration of almost 5 times faster when using the SDK compared to a Core i7.

TMPGEnc is known for delivering better IQ at any given bitrate, so I'm really curious to see if they compromised the IQ when implementing their CUDA and Quicksync enabled transcoding algorithms (like everyone else did) or if they actually implemented them to deliver comparable IQ at faster transcode times.

 

[Johnnie.it]

Idontcare: alright, this is all sweet and beautiful, but judging from what info I gathered around the internet, CUDA and quicsync are still useless stuff, since all I've got is an ATI 6950 and (hopefully, in a few days) an FX-8150.
From what I understood, I need avivo software.

I also found out that, apparently, there's a build of Handbrake that uses openCL (that should be platform agnostic).
Unfortunately, it's still an alpha release so it may take some time to have a working version.

This is the most interesting link I found yet: http://www.mainconcept.com/products/sdks/gpu-acceleration/opencltm-h264avc.html
Unfortunately, there's not a link where I can download a software for that.
Avivo (http://www.amd.com/us/products/technologies/ati-avivo-hd/Pages/ati-avivo-hd.aspx) seems interesting too, but support apparently stops at 4000 series (on some pages it says 4000 and up).

This is all very confusing...

 

[lopri]

Not to discount its merits, but TMPGEnc wsn't anything special for me. It looks like it earned its name during the days of DVD-ripping. I am unsure whether it's the right tool for the OP.

When I tried out TMPGEnc ver.4, and the thing was quite buggy. Wasn't impressed with its presets (i.e. put out corrupt images, de-syncs, missing audio, etc.) and CUDA did not work at all. I say that because the check box (checked/unchecked) did not do anything. I had a GTX 580 at the time. They did advertise that it works with certain custom post-processing but I did not bother to use those because my general goal when transcoding is to get the picture quality as close to as the originals while reducing the file size.

Their PR was kind of amusing as well.

http://tmpgenc.pegasys-inc.com/en/product/te4xp_new.html (Scroll down to "Test System Configuration")

Look at each test and what they do. The first test (deinterlacing and resizing) is what one might call "encoding" or "transcoding," and it shows Cuda (GTX 280) being half the speed of CPU transcoding (Q9450). Then their subsequent tests are cosmetic post-processings one at a time. Not only had I no idea what "Smartsharp" does, but if it's anything like sharpen filters that are found in many media players (such as MPC-HC), those could be done in real time with minimal computing resources.

But anyhow, I did not try those ambiguous filters because they often do more harm than good for image quality. Especially with labels such as "Smartsharp," "Sharpening Edges," it is hard to know what you're going to end up seeing after it's done. And if it turns out you don't like it.. tada. You've just wasted your time and energy. A better practice, IMO, is to apply only proven, precise tweaks and leave the rest as close to the original.

So I ended up just trying it out, without buying. There was simply nothing I could not achieve just as well or better from (free) alternatives.

I meant to give another shot at its version 5, which now is called "TMPGEnc Video Mastering Works," but upon learning its draconian DRM scheme, even for trial version, I decided not to. From what I can see, the product's virtue is that it wants to be the encyclopedia of every multimedia format known to man. And so long as things do work as they advertise, I suppose there is a value to having an all-in-one package like that. For instance, it might save you time if you run into an unexpected device that takes older formats.

With filters like these..

But a more likely scenario for personal use is that that you don't need all that. You probably need only a few for your needs (portables, archives, and perhaps cameras) And for such purposes there are lots of cheaper/free alternatives, and they work better. In my case, Handbrake was just that.

Don't get me wrong. I do hope the days come quick where GPU is able to beat CPU in transcoding. But we are not there yet.

A look at hardware video transcoding on the PC

OP, you mentioned using x264 for encoding. What tool are you using? Adding a bunch of post-processing filters isn't necessarily going to improve the video quality unless you know what's exactly being done.

If you want a decent quality for backup'ing 1080p materials, there are tons of cheap apps out there. (including Handbrake, which is free!) If you want to be as real as it gets, you could try tools like these:

HDConvertToX
MeGUI

There are some learning curves for these so it's your decision whether it's worth the time/effort.

 

[lopri]

P.S. Mainconcept is a codec development house. They sell their code to 3rd parties, not directly to consumers. So when you buy Sony Vegas or Adobe Premiere Mainconcept codecs come with them. And Avivo is.. so 2006. I don't know whether AMD still uses that brand, but it's now referred to as UVD. (Not exactly the same thing, but does the same thing ^^ )

 

[Johnnie.it]

OP, you mentioned using x264 for encoding. What tool are you using? Adding a bunch of post-processing filters isn't necessarily going to improve the video quality unless you know what's exactly being done.

I've found that some filters really help indeed. Deblock, to name just one, really lets me push down the bitrate quite a lot (I'm still talking about 8.5-10+mbps for 1080p, depending on the lenght of the movie and, of course, its original quality

If you want a decent quality for backup'ing 1080p materials, there are tons of cheap apps out there. (including Handbrake, which is free!)

I'm actually using handbrake...

If you want to be as real as it gets, you could try tools like these:

HDConvertToX
MeGUI

There are some learning curves for these so it's your decision whether it's worth the time/effort.

I don't know...I mean, as long as it doesn't take weeks of experimenting, I usually find it's worth it...
Besides, doom9 has always been aimed at dvd, I've always had a hard time trying to understand what tools can be used with h264 and how. Well...truth be said, for that there's lots of guides, but they still explain how to do everything with dvds. I can still get them to work with blurays, but that will be a lot of effort I don't really need.
Handbrake really gets as is as it can, if I want. After a bit of experimenting, some reading around and a couple of test runs I can get the perfect mkv file, so...I'd rather stick with that for the time being, unless I can find a major improvement (like gpu assisted encoding).
I was already looking into MeGUI anyway.

P.S. Mainconcept is a codec development house. They sell their code to 3rd parties, not directly to consumers. So when you buy Sony Vegas or Adobe Premiere Mainconcept codecs come with them.

Yes, yes...I just mentioned that page because I had no idea about OpenCL.
That sounds brilliant.

And Avivo is.. so 2006. I don't know whether AMD still uses that brand, but it's now referred to as UVD. (Not exactly the same thing, but does the same thing ^^ )

Well apparently now avivo is part of the amd codec media pack, which I just downloaded, only to learn it was already part of the catalyst suite I got installed. Well I updated everything as well.
What baffles me is that I'm supposed to have a button to launch the video converter from the catalyst control center (well, now it's AMD VISION engine control center) but it's missing.
I also found this thread:
http://www.tomshardware.co.uk/forum/344154-15-start-video-converter-driver
but none of the suggested solutions worked for me.
I'll look deeper into that.

Anyway, besides this codec media pack, I just found out about AMD EVC engine (but apparently it's a bit too soon yet), this article here on Anandtech What We've Been Waiting For: Testing OpenCL Accelerated Handbrake with AMD's Trinity and some news in a FAQ for handbrake:

Why doesn't HandBrake use Grand Central Dispatch or OpenCL?

First off, HandBrake is cross-platform. Until GCD's library is a standard component in Linux distributions, and OpenCL SDKs stabilize, these features are mostly limited to OS X.

Second, HandBrake is already fully multithreaded. While Grand Central makes it easy to thread an app that runs on a single CPU, it has no real benefit for an app that was already threaded, the hard way.

Third, OpenCL's benefits are limited for dealing with video. Unless every single step of the encoding process happens on the GPU, data must constantly be shuffled from the host's RAM to the video card's RAM, and them from the global VRAM to the local and private VRAM, and then back the other way after it's been processed. This goes over the PCI bus, and is a huge bottleneck for the encoding process.

I'd just hate to have to buy a mac for that. I'm not really a mac fanboy...well let's say that, right or wrong may it be, I kinda despise everything apple.

Besides, I downloaded a trial version for this Wondershare Video Converter Ultimate because it seemed to support gpu encoding.
Well, it seems some useless junk to me.
It pretty much just lets you choose resolution and bitrate, and that's about it.
No multiple audio tracks supported, of course it forces an ac3 encoding, there's no way to have a simple passthrough (I usually keep dts tracks), and only one subtitle track.

I'm just going for a test run against handbrake, but I have to say I'm not impressend in the least, so far.

 

[anongineer]

I don't think the HandBrake FAQ is recommending that you get a Mac. If anything, it's saying that certain features offered by a Mac are not helpful and so HandBrake's developers see no point in using them.

CPU encoding is still the way to go. The HandBrake FAQ's comments on OpenCL illustrate why there's not much to be gained from GPU assisted encoding. The data gathering stage is great for a GPU, but the analysis and decision making are much better suited to a CPU. Going back and forth incurs latency. An APU may cut down on that latency, but they won't have as much number crunching power as a discrete setup.

 

[Johnnie.it]

I don't think the HandBrake FAQ is recommending that you get a Mac. If anything, it's saying that certain features offered by a Mac are not helpful and so HandBrake's developers see no point in using them.

Of course it doesn't, that would be just ridiculous.
But they're saying that openCL implementations are way better on mac, that's why they're already on osx

CPU encoding is still the way to go. The HandBrake FAQ's comments on OpenCL illustrate why there's not much to be gained from GPU assisted encoding. The data gathering stage is great for a GPU, but the analysis and decision making are much better suited to a CPU. Going back and forth incurs latency. An APU may cut down on that latency, but they won't have as much number crunching power as a discrete setup.

Well, if I could just lose a couple of hours for the first pass (which is all read-ahead, which a GPU can do way better than a CPU) I'd be already happier

 

[Idontcare]

I've found that some filters really help indeed. Deblock, to name just one, really lets me push down the bitrate quite a lot (I'm still talking about 8.5-10+mbps for 1080p, depending on the lenght of the movie and, of course, its original quality

So true. Any filter that removes what is essentially pixel noise, either within frame (spatial noise) or frame-to-frame (temporal noise), will enable better better compression (lower bitrates) with the same IQ.

There is a huge benefit to be had in reducing noise. But you do have to be careful, if you make the filter too aggressive then not only does it filter out the noise but it will also start to filter out the intended color gradients and motion in the video itself.

You really have to pay attention, but a little noise filtering goes a long ways.

What you don't want is to have pixel noise and have your encoding algorithm expending all of its bitrate towards faithfully reproducing the noise instead of faithfully reproducing the edges and motion in the movie. Remove the noise and let the encoding expend its bitrate budget to reproduce the stuff you want to see in the final product.

 

[synapse46]

If you have an am3+ board the 8350 offers up good performance.

http://www.techspot.com/review/586-amd-fx-8350-fx-6300/page5.html

 

[Deaks2]

I'll eat my words, 8350 is a good deal for video encoders if you have cheap electricity.

 

[synapse46]

I'll eat my words, 8350 is a good deal for video encoders if you have cheap electricity.

Turn the lights off when you're encoding, problem solved. If you paid $400-1000 for your CPU, you may not be so concerned.

 

[inf64]

One poster over at SA calculated how much $$$ you may save with lower power draw on the CPU side and it's trivial to say the best:

I really do hate this argument. Even if you ran your FX at 100% load for 10 hours a day 7 days a week and it pulled 100 more watts than the competition, it would still only be $40 a year at 10 cents a kilowatt hour. Considering 3770k costs $120 more, it would take 3 years of running your computer at 100% for 10 hours a day every day to break even with the money you'd save by going 3770k for power consumption.

This argument really needs to die. It makes people look completely inept. There are plenty of places where Piledriver has short falls, but whining about $10 a year in extra power usage (realistically, how often is your CPU going to be running full load? I would guess a fourth of 10 hours a day every day or less) on a product that falls between a product that cost $120 and $30 more is ridiculous.

Unless you live in some remote area where you're paying 50 cents a kilowatt hour or you plan on buying 1000 FX CPUs, it doesn't matter. If FX cost the same as 3770k, power consumption would matter. It's why it mattered with Fermi so much and you don't hear me whining about it with FX.

So to sum up: it's a bad argument even if you run your PC at 100% load 10/7 . Unless you live in a country that has obscenely expensive electrical energy,which is rare.

 

[Deaks2]

The OP is from Europe. $0.10/kwh does not exist there... As mentioed earlier in this thread, in Denmark electricity costs $0.40/kwh, in the UK it is about $0.25/kwh.

 

[inf64]

Ok so it would take less than a year in Denmark to chew thru the price advantage FX has but only if the OP would to be running his PC 10 hours under full load every single day. Highly unlikely usage scenario don't you think?
For UK under same scenario it would take approx. 1.2 years or 440 days. Again 10 hours at 100% load every single day.

At any moment the user is not running his PC at 100% load the power discrepancy drops to almost nothing since both platforms have near identical idle power draw. If he runs one threaded workload the power draw difference is much smaller as xbitlabs measured(114W for 8350 and 88W for 3770K-a mere 26W). Also note that the original calculation @ SA topic is based on 100W difference and xbitlabs measured 73W under 100% CPU load which is much lower.

 

[AtenRa]

I will insist on waiting for PileDriver, FX8320 will cost less than current FX8150 and have the same performance + lower power (dont ask me to provide evidence )

From the egg,

FX8150 = $189.99
FX8320 = $179,99 ($11 price gouging over MSRP of $169)

I hope you didnt buy the FX8150

 

[Durvelle27]

^^^ i agree


http://www.youtube.com/watch?v=vzkUvkXfGKw&feature=g-all-u

 

[synapse46]

So how did the tester get the 8150 to pull 199 watts, is that accurate?

Also, who does two pass encoding?