encoding speed

[Tuptim]

Hi,

I'm a long time DVD encoder that are now moving to Blu-Ray, i try to go for best quality always and foremost. That why i used CCE 6 pass when encoding for DVD and not DVDShrink (amazing app) or other less quality encoding tools.

Now i've a speed problem, old computer (AMD Atlhon 64 X2 Dual Core 3800+ | 4gigs DDR2 Mem), this takes me about 7 hours for a DVD encode and i'm fine with that (DVD to DVD).

But with MeGui veryslow preset (Blu-ray to Blu-ray) it takes me around 5 days (Automated 2-Pass). I've read some benchmarks results and some say Intel i5 750 are better than AMD Phenom II x4 965 BE, others say the opposite. I dont have the money to buy a Core i7.

So what i'm asking you guys is for some real-life encoding speed's with a similar rig (i5 750 4 Gig DDR3 / Phenom II x4 965 4 Gig DDR3), please bear in mind that i barely have the money to buy the Core i5, The AMD is €50 cheaper and i'm only buying this for encoding (im perfectly fine with gaming and general purpose application as i am) and that i pretend only to encode Blu-Ray to Blu- Ray, Not Blu-Ray to .MKV or 1080 to 720, and using almost the max settings for quality.

Thanks in advance for your comments

[nurbs]

The two processors have about the same encoding speed. See here (http://www.anandtech.com/show/2832/12). Second pass results are usually more interesting.

[stax76]

Looks like the AMD one is faster as well as cheaper, anybody surprised?

[Tuptim]

Thanks for the additional info.
Based on that 2nd pass graph a content with 177192 frames (2h:3m) would take around 2.3 hours.
So its fair to assume that it is with Medium, Very Fast, Ultra Fast presets ?

[nurbs]

Looks like the AMD one is faster as well as cheaper, anybody surprised?

I'm not disputing that the AMD system is the better choice, but while it's 6% faster in first pass the Intel system is 3% faster on second pass which for most scenarios means Intel is slightly faster faster. The benchmark is usually made with Intel Turbo Mode disabled, which probably has a negative influence on the first pass results.


Thanks for the additional info.
Based on that 2nd pass graph a content with 177192 frames (2h:3m) would take around 2.3 hours.
So its fair to assume that it is with Medium, Very Fast, Ultra Fast presets ?
The x264 revision used for that benchmark is 1342. They use settings similar to the Medium preset. I think the speed would be about the same.

[Tuptim]

You guys are great, just one more thing.

Based on Nurbs prediction that the benchmark results are on medium and that a 2h content would take 2.3 hours can someone give me an extimate of how much time the same content would take in the veryslow preset (either in time or fps).

Thanks

[mp3dom]

Core i7 920, a film of 140min. (1080p24) encoded at 30 Mbps in about 15-16 hrs with preset veryslow (2 pass vbr with fast firstpass).
Core i7 980, PAL SD encodes at about 15-16 fps at 3 Mbps

[nurbs]

Depends on bitrate and resolution. Assuming 720p at 4 Mbps I'd say about 5 fps on second pass, 480p at about 1 Mbps around 15 to 20 fps. This is from my own experience using the chart to convert my encoding time to the processors you listed and might be way off.

edit:
I just remembered that --preset veryslow uses 8 b-frames with --b-adapt 2, so first pass will be slow.

[Tuptim]

I was going more for the mp3dom setting.
Bitrate above 20 Mbps 1080p content

So i will assume with an i5 750 / Phenom 2 x4 965 it would take 20-24 hours. I think i can take that, compared to the 5-6 days it take with the rig i've atm.

Thanks you guys for all the information you shared.

PS: I will try to post my results here when i buy the new stuff and make an encode (it may take some time)

[Groucho2004]

You should really consider the AMD 6 core 1055. Same price (or cheaper) than the i750 and a lot faster with x264 encoding.

Have a look at this (http://forum.doom9.org/showthread.php?t=155791) thread.

[7ekno]

+1 for the x6, makes a world of difference for very little outlay (and is very easily over-clockable to 4GHz on air cooling!) ...

Has saved me $$ and plenty of time !

Buy a NVidia GT240 GPU and with DGDecNV you offload decoding (deinterlacing, cropping and resizing) to GPU ;)

GT240 + DGDecNV + x6 1055T o/c to 4060MHz (14x290) and your over 110fps in the benchmark from the other thread ;)

7ek

[nurbs]

Buy a NVidia GT240 GPU and with DGDecNV you offload decoding (deinterlacing, cropping and resizing) to GPU ;)
Have one of those. The 2xx series is a good choice for that, especially if you have to deal with 1080i VC1, like for instance a some of the BBC Blu-Ray discs. Very inexpensive as well.

[Tuptim]

Thanks Groucho2004, 7ekno, nurbs.

My funds are limited, here the prices are:

AMD PHENOM II X4 965 BE Quad-Core 64Bits 8MB BOX SKAM3 - €174
Intel Core i5 750 2.66GHz 8MB BOX - €191.90
AMD PHENOM II X6 1055T 2.8GHz 9MB BOX SKAM3 - €199.90

The difference dosent seem much, and it isnt, but remember that i need to buy a new Motherboard and new DDR 3 RAM too, to fit this new stuff.

I appreciate the NVidia GT240 GPU tip but i dont intend to do any deinterlacing, cropping and resizing, buying a new Graphic Card its completly out of budget. And I HATE NVidia Video-Output (that i use alot) so i will stick with my faithfull ATI.

I think i can get more €26 to buy the 1055T.

Thanks alot for the heads up on that.

[stax76]

buying a new Graphic Card its completly out of budget

For DGDecNV a 10-20€ eBay card would be sufficient.

[gahz]

here's a better cpu comparison from techarp http://www.techarp.com/showarticle.aspx?artno=669&pgno=3

[Assassinator]

here's a better cpu comparison from techarp http://www.techarp.com/showarticle.aspx?artno=669&pgno=3

No, actually, that a very bad comparison. It's a very bad reflection of actual encoding performance, and is very very deceiving.

The CPUs are ranked based on a simple arithmetic average of the speeds of the 2 passes. This is dumb, because fps is a measure of rate (speed/time), and it's extremely poor mathematics to simply average rates like that.




Lets use an easy example to try to illustrate why using an arithmetic average for this kind of stuff is bad...

lets assume...
CPU A: 1st pass fps = 90. 2nd pass fps = 10.
CPU B: 1st pass fps = 50. 2nd pass fps = 50.

Arithmetic average...
A = (90+10)/2 = 50 frames/second
B = (50+50)/2 = 50 frames/second

That chart would list these 2 CPUs as being equivalent. Lets see the truth.

CPU A total encoding time, for a clip of 1000 frames.
= first pass time + 2nd pass time
= (1000/90) + (1000/10)
= 111.1111.... seconds

CPU B total encoding time
= (1000/50) + (1000/50)
= 40 seconds


CPU B has almost 3x the actual performance of CPU A. Would you say they are equal, as the arithmetic mean predicts them to be, both 50fps?
- I know I wouldn't, CPU A has 50fps my ass. Only really (1000/111.111...)*2 = 18fps average.





Now you're probably saying.. blah blah, extreme examples, who cares about those. The numbers in the charts aren't nearly as different as 90-10 and 50-50... The chart would have you believe a Q6700 @ 2.66GHz is faster than an i7 860 @ 2.8GHz... while if you look at the 2nd pass results alone, the 860 scores 25.89fps, while the Q6700 is a measly 17.27fps. That says everything about how good a measure the arithmetic average really is, for rates like this.



Then you may ask... what would be a good way to average the numbers for this kind of thing?
- harmonic mean (http://en.wikipedia.org/wiki/Harmonic_mean).

[Didée]

Also, to note it once more, the Nehalem numbers are somewhat skewed.
Please remember to turn off Intel Turbo Mode in the BIOS for accurate results with Intel Core i7 or Core i5 processors!
In real-world @ stock settings, an i870 "2.93 GHz" in fact is running at 3.2 GHz due to Turbo (when under full load, as e.g. in an x264-2nd'pass...) But they say this is not accurate.

It seems that to TechARP, "accurate" means to give all Nehalem CPUs a penalty of 5~10%. Which is not little when numbers are close.

[aegisofrime]

I would say that with Sandy Bridge and Bulldozer nearing release, us people looking to upgrade for better encoding performance should hedge our bets. AVX should be a major game-changer, on top of whatever architectural improvements each CPU implements.

[Kuukunen]

But with MeGui veryslow preset (Blu-ray to Blu-ray) it takes me around 5 days (Automated 2-Pass).
By the way, correct me if I'm wrong, but last time I checked, MeGUI's presets were horribly outdated. I'd suggest you use x264 from the commandline and its built-in presets.

Also, if you're fine with that many hours of encoding, offloading the decoding part to GPU doesn't mean THAT much more speed.

[Sharktooth]

@Kuukunen: they are no longer "horribly" outdated. Megui supports x264 presets.

[Kuukunen]

Thanks for the correction. :P

[board123]

Also, to note it once more, the Nehalem numbers are somewhat skewed.

In real-world @ stock settings, an i870 "2.93 GHz" in fact is running at 3.2 GHz due to Turbo (when under full load, as e.g. in an x264-2nd'pass...) But they say this is not accurate.

It seems that to TechARP, "accurate" means to give all Nehalem CPUs a penalty of 5~10%. Which is not little when numbers are close.
I thought Turbo only turns on when the load is high and not all available cores/threads are used. x264 is uses all available threads, so Turbo wouldn't apply here.

[AlekseiV]

Now i've a speed problem, old computer (AMD Atlhon 64 X2 Dual Core 3800+ | 4gigs DDR2 Mem), this takes me about 7 hours for a DVD encode and i'm fine with that (DVD to DVD).

But with MeGui veryslow preset (Blu-ray to Blu-ray) it takes me around 5 days (Automated 2-Pass). I've read some benchmarks results and some say Intel i5 750 are better than AMD Phenom II x4 965 BE, others say the opposite. I dont have the money to buy a Core i7.Then why are you using the veryslow preset if it takes so long? Choose a preset that is fast enough.

[sneaker_ger]

I thought Turbo only turns on when the load is high and not all available cores/threads are used. x264 is uses all available threads, so Turbo wouldn't apply here.

No, it will also turn on if all cores are used. It will just not go as high as with fewer active cores. For example my i7-860 (sold as 2,8 GHz) will go to 2,93 GHz if fully used by x264. It will go up to 3,3x GHz for single core use. (wiki (http://en.wikipedia.org/wiki/Turbo_Boost))

[Didée]

No. Turbo states on Nehalem are 3-fold. 1st state kicks in even when all cores are under full load. 2nd and 3rd turbo state can only be used while other cores are parked in a C-state.

Example i870:

nominal multi: x 22
4 cores loaded: x 24
2 cores loaded: x 26
1 core loaded: x 27

At true stock settings, x264 using plenty threads will be driven by 133 x 24 =~ 3.2 GHz.
With turbo disabled as per TechARP, x264 will be driven by only 133 x 22 =~ 2.9 GHz.