Good bang for buck cpu for x264 encoding

[nikusor665]

What is right now a good bang for buck cpu for x264 encoding ?
I'm thinking fx-6100 or fx-8120 right ?

I have a core2duo and needless to say I want an upgrade.

[hajj_3]

AMD are cheaper but they are quite a lot slower at x264 encoding at 1st pass. The Core i5-3570K ivy bridge is a good processor. http://www.xbitlabs.com/articles/cpu/display/core-i7-3770k-i5-3570k_7.html

[nm]

AMD Bulldozer is pretty ok for the price. However, if you're encoding 24/7 for two or three years, they might cost more than comparable Sandy/Ivy Bridge CPUs in the end because of higher power consumption and electricity costs. Unless you don't have to pay the electricity bill or can use the heat for something. :)

[hajj_3]

AMD vishera based processors will be out in Q3 btw so within 4 months ish, those will increase the performance, it is difficult to say by how much. There will be 1 or more 10 core AMD vishera processors.

[Atak_Snajpera]

I'm thinking fx-6100 or fx-8120 right ?
don't even think about fx-6100.

results from x264 FHD Benchmark. More results (http://www.x264fhdbenchmark.republika.pl/Results.txt)
17.8 fps - AMD FX-8120 @ 3.1GHz ( 8C / 8T ) MMX2 SSE2Fast SSSE3 FastShuffle SSE4.2 AVX XOP FMA4 SSEMisalign LZCNT
16.9 fps - AMD FX-6100 @ 4.2GHz ( 6C / 6T ) MMX2 SSE2Fast SSSE3 FastShuffle SSE4.2 AVX XOP FMA4 SSEMisalign LZCNT
16.8 fps - AMD Phenom II X6 1100T @ 3.3GHz ( 6C / 6T ) MMX2 SSE2Fast FastShuffle SSEMisalign LZCNT

Even old x6 is only a little bit slower than new FX "8-core" :)

[nikusor665]

Thanks for that link, really useful.

[mandarinka]

It think it would be easier to use if it wasn't littered with overclocked results. OTOH, there is no stock 8150 for example... :(

[Atak_Snajpera]

stock 8150 will have something around 19-20 fps

[mandarinka]

Scaling the 8120 by frequency, one gets something over 20.5 (in x264 benchmark 4.0 it was close to i7 2600 IIRC).

Anyway it is true that the Intel cpus can recover some of the price premium back if they are run at full load a lot. But amusingly x264 is almost the best-case scenario of Bulldozer's performance, heh.

[Balthazar2k4]

Best bang for the buck IMO are the 6-core Intel i7 processors. I have both a Gulftown i7-980x and the new Sandy Bridge-E i7-3930k and they are ridiculousy fast with x264. I can encode a movie using "Slower, Film" preset in realtime on the 3930k. Now with RipBot supporting distributed encoding I will run them in tandem and use the "Very Slow" setting.

If you can afford one, I would say go for it.

[LoRd_MuldeR]

Best bang for the buck IMO are the 6-core Intel i7 processors. I have both a Gulftown i7-980x and the new Sandy Bridge-E i7-3930k and they are ridiculousy fast with x264. I can encode a movie using "Slower, Film" preset in realtime on the 3930k. Now with RipBot supporting distributed encoding I will run them in tandem and use the "Very Slow" setting.

You are confusing "maximum performance per money" with "maximum performance" ;)

I think nobody will doubt that the Core i7 Hexacores currently provide the maximum performance you can get from a single-socket CPU.

Nonetheless the performance doesn't legitimate the extraordinary price of these CPU's. As always, for such "high end" models, you pay a lot of extra money for relatively few extra performance.

So if you want to get the maximum Intel performance for your money, you probably don't want to go above a decent i5 Quadcore at the moment...

(Instead of spending twice the money for a "high end" CPU, I would rather save that money for the next upgrade)

[Didée]

Well, you can look at this from many different angles. Another point is the ratio (computing.performance)/(energy.consumption). How do the actual hexacores fair in comparison to quadcores? (Serious question, I'm not uptodate with the sandy generation.)

Yet another point is raw time, and how valuable the saved time is to you. Paying like 500 bucks extra for a CPU surely is not trivial. But then, if you use it a lot, and it saves you 50 or 100 hours each month, perhaps that extra is well spent.

It's difficult to come to general conclusions, when the possibly-relevant facettes are so many.

[Balthazar2k4]

I can understand how many (most) people would see Intel's hexacore models as being spendy. They are. That said, when I found myself tasked with re-encoding my entire BD collection (>400), the speed of these processors suddenly become invaluable. The ability to knock out 40+ movies in 1080p a week was simply worth the extra couple hundred dollars over a quadcore. Just my perspective.

[Guest]

We're not discussing the meaning of "best" are we? Sort of a meta-violation, if you will. No problem, carry on. :p

[SubOne]

FX-8120 is only $169 and at that price it is surely a lot of bang for the buck. 17.9 in the FHD benchmark. The closest priced Intel is Core i5 2400, which goes for $189 but is only good for 14.3 fps. Heck, it's on par with the new Core i5 3570K, which goes for $239! AMD motherboards are cheaper too. On the other hand, FX has a TDP of 125W while 3570K is 77W and 2400 is 95W. Of course, those Intels are way better at single thread / lightly threaded stuff like gaming.

[mandarinka]

"That said, when I found myself tasked with re-encoding my entire BD collection (>400), the speed of these processors suddenly become invaluable."

You make it sound as if somebody was forcing you to do it and it was some sort of a vital action, lol...

[benwaggoner]

And bear in mind that the cost of the CPU is only one part of the system. The chassis, motherboard, display, RAM, storage, etcetera are going to be more fixed. Even if a CPU gives 2x the performance at 3x the price for the CPU, it could still offer better price/performance for the whole system.

And power draw and cooling costs can be a significant factor as well.

My new workstation has dual six-core X5650s and 9 GB of RAID, and well worth the money considering how much faster I can turn around content.

[mandarinka]

"Improving" the price/performance ratio by including the whole system is ridiculous. If some way of calculation lessens the differences between things, it isn't a good way of meassuring stuff.

This could matter for example for different platforms, if for one, you had to use more expensive memory (rambus), or the boards were much more expensive. Then you can take that extra cost into consideration. But equipping the systems with needlessly fancy boards for example, to make the cpus look less expensive for their relative worth, that's simply misleading. And precisely the same thing happens when one adds prices of all components into the equation. Doing this for case, speakers and display is silly - those are fixed costs that are same for both, they have no place in comparison.

To drag this into complete absurdity, why shouldn't I include my hi-end headphones and that atomic clock-powered preamp I got for cheap?
When I do that, it seems the cpus are almost free, so why not get both and drive with my car over one of them.

[benwaggoner]

"Improving" the price/performance ratio by including the whole system is ridiculous. If some way of calculation lessens the differences between things, it isn't a good way of meassuring stuff.
The sensible metrics to me here seem to be the capital cost per unit of encoding throughput, or better yet, TCO per unit. In that, the comparison needs to be between the total cost of an encoding system to purchase and operate.

An encoding system doesn't need to have headphones, speakers, or monitors, but it is sure more than a CPU! The specifics will vary depending on architecture. Building out a farm of dozens of encoding systems with shared storage is very different than a couple of boxes working off their own internal storage.

In my experience pricing these things out (going back to the PowerMac 8100/80, mind you), going up in number of cores up to two sockets generally pays off with today's multithreaded, multi-bitrate encoders. But going for the extreme high clock speeds rarely does, since that last 20% of performance can be another $1000/CPU. More than two sockets also doesn't pay off since processors get more expensive yet, and maximum clock speeds keep going down.

Of course, if you have minimum throughput requirements (like being able to turn around a 150 minute movie in 10 hours at multiple bitrates), "fast enough" winds up as a second critical factor to TCO per unit encoded.

[mandarinka]

I'm not saying that these factors are irrelevant. They do matter, but if we are talking about the bang-for-buckness of the actual CPU, they need to be mostly ignored. Sure, if the cheapest usable ram/board/psu/cooler combination for system B costs more than for system A (or if you need to install a gpu for system A but not system B), you need to add that difference. But meassuring the actual CPU should be just that.

When making the actual box, you are already going into another problem, because the system can be configured for variety of uses (gaming, just encoding, reliability with raid/ups etc). But that would mean that as you reconfigure the system for another task, the CPU's price/performance ratio changes? I think we can agree that thinking like that would be absurd.

I believe you are actually thinking of this second level of problem - of making the whole box. Evaluation on this level can change the cpu's worth (for example - FX supports ECC RAM, you need to get Xeon for that going with Intel and motherboard will also get more expensive). But when we are talking the general worth of the cpu for x264 encoding, it is clear that we can't include the whole system, because we don't know what configuration to pick for that.

For the first level of the problem, for general evaluating of the price/performance ratio, we can't use the whole system price at all. As I said, we should add in the delta in prices of minimal configurations, but that's it (example: if the cheapest satisfactory mobo for system A is 80$ and for B it's 100$; but B is okay with 10$ cheaper cooler and I can use the same components for both systems otherwise, then I will adjust the price of system B by +10$, for the comparison). If I calculate anything more into the prices, I am misleading the results.

BTW I think that when picking between the three mainstream desktop platforms (socket 1155, socket fm1 and socket am3+), the prices of motherboards are pretty close, one can drive all them with a simple PSU and RAM is the same. AM3+ can be got with IGP chipset for cheap. That means that the encoding performance relative to cost of the resulting boxes is determined nearly solely by the cpu's price. If I get i7-2600, I get over twice the performance of i3-2100, for more than twice the price. If I get FX-8150, I gt something under 2600's perf for somewhat lower price than 2600. It's completely fair to judge it like this imho. You can then wonder, if the money put into purchase of lower-powered but more expensive processor would return in reasonable time...

[Blue_MiSfit]

@benwaggoner:

I typically follow a similar train of thought when specing encoding systems. My new standard is 2 sockets per system, populated with 6 core Xeon E5-2430 CPUs, in either a 1U or half-height blade form factor. These have incredible performance, and are surprisingly inexpensive. Adding more clock speed doesn't seem to usually be cost effective nearly as much as scaling out, for most applications.

Occasionally I use encoding workflows that don't multi-thread well, and where doing more encodes in parallel isn't the answer, and some video analysis tools I use are licensed on a per-core basis (at a very high cost), so there it makes sense to use higher clocked parts.

Derek

[bikar]

I find Fx-8320 best cpu for x264

we did some real tests here (http://www.bug.hr/forum/topic/benchmark/x264exe-benchmark-2/169095.aspx)

http://www.bug.hr/_cache/94987685008b5ca025217325fd8ac5ca.jpg
http://www.bug.hr/_cache/cd0cebb2955755d74c765125dbb218f8.jpg
http://www.bug.hr/_cache/67e129e6659fbd6a7a8035520cef7ced.jpg?rand=950672582

[mandarinka]

FX-8320 and FX-8350 aren't bad value (ideally though, the 8320 would be 95W) when you have encoding via x264 in your mind, because that is almost the ideal scenario for their use/benchmarking.

More importantly, as you can see, AMD has much more cpu capabilities. While the Sandy-Bridge based Core i7 has 6 capabilities, AMD's Piledriver has whopping 13. That's more than double, people! :sly:

[Groucho2004]

FX-8320 and FX-8350 aren't bad value (ideally though, the 8320 would be 95W) when you have encoding via x264 in your mind, because that is almost the ideal scenario for their use/benchmarking.

More importantly, as you can see, AMD has much more cpu capabilities. While the Sandy-Bridge based Core i7 has 6 capabilities, AMD's Piledriver has whopping 13. That's more than double, people! :sly:

If you look at the power consumption figures in various tests (here (http://www.xbitlabs.com/articles/cpu/display/amd-fx-8350_8.html) for example) it becomes clear that the "better bang for the buck" argument becomes invalid, especially in scenarios like video encoding on a daily basis.

Also, all the additional CPU capabilities don't help if they don't translate into more performance.

The bottom line is that even the latest Piledriver core consumes almost twice the power at the same performance level as Ivy Bridge or Sandy Bridge E.

Most people claiming that the AMD processors are better value seem to still live with their mommy and daddy and don't have to worry about electricity bills.

[Atak_Snajpera]

keep in mind that people in eastern europe don't earn 1200 - 1600 euro monthly. if you are lucky you will get 300 - 400 euro. also average joe probably won't be encoding 24/7.

[mandarinka]

What he said.

The electricity problem depends on how much are you going to run at 100% load (anecdotally, I'd say I have full load less than 20 % of runtime, plus obviously the PC isn't on 24/7). In idle, the power consumption is on par (the only differences come from the motherboards you choose). There are tradeoffs you need to make, it's not as simple as your mom/dad 'wisdom' (@ Groucho2004).

And I do think you failed to spot an ironic smiley at some point too :)

[bikar]

here in croatia Fx -8320 is 1300kn (224$) and i7 2600k is 2300kn (393$) + more expensive MBO for intel cpu

I think it's pretty clear for avarage (poor) Ivan what to do :D

[benwaggoner]

This really requires a full TCO estimate including purchase price and operating cost. Pixels-per-picojoule can wind up being a bigger total factor than purchase price.

Throughput and latency are also different targets. Capital costs go up as maximum turnaround time goes down, since that needs lots of fast cores, likely with two sockets. However, operating costs may vary less. Dual-socket systems with lots of cores but a somewhat slower clock speed can offer better throughput and latency and lower power costs than a few very fast cores. Capital cost is going to be a lot higher, but that can wind up being minor compared to operating cost at heavy loads.

Workflow architecture also matters enormously. If encoding 16 different bitrates from a single source, a frameserver system that requires only a single decode and source filter pass is a lot more efficient than running AVISynth 16 different times (or 32 for 2-pass).

For 24/7 multiple bitrate encoding, two-socket Xeon systems with the max number of cores per socket have long been a sweet spot for TCO.

Judicious setting tuning and build selection can also matter significantly. Just going to 64-bit builds instead of 32-bit is a nice boost. Using a single analysis for multiple output bitrates is a big savings. And not using "Placebo" is a huge savings :)!

Net-net, "best bang for the buck" can vary a lot depending on the flavor of bang being sought.

[ajp_anton]

In colder places, using more electricity during the winter isn't much of a problem.

[aegisofrime]

One point I would like to add is whether one is doing just pure transcoding/encoding, or doing some postprocessing work like with Avisynth along with it. AMD has worse IPC than Intel, and in some cases multi-threaded Avisynth is unstable. In such cases, the superior core count of Piledriver is irrelevant.

[ajp_anton]

IPC is irrelevant if you can compensate with frequency (which AMD can't really do anyway). What you wanted to say is AMD has worse single-threaded performance.
Hopefully Vapoursynth will become popular as it seems to be much better at multithreading.

[nibus]

All the x264 benchmarks I've seen comparing the FX series with Ivy Bridge only run a 2-pass test, such as this one:

http://www.anandtech.com/show/6396/the-vishera-review-amd-fx8350-fx8320-fx6300-and-fx4300-tested/3

Where the 8350 comes out slightly faster on 2nd pass, but slower on first.

I'm wondering how this could translate to CRF encoding? I rarely use 2-pass anymore.

[Groucho2004]

I'm wondering how this could translate to CRF encoding?
In my experience, CRF and the second pass of a 2-pass encode at roughly the same speed using the same settings.

[ajp_anton]

But what about the poorly threaded 1st pass things that CRF has to deal with? How well does CRF max out 8 cores (with a "reasonable" preset)?

[Blue_MiSfit]

SD? You probably won't - depending on what's doing the decoding.

HD? Should be no problem...

[Juce]

http://www.anandtech.com/show/6396/the-vishera-review-amd-fx8350-fx8320-fx6300-and-fx4300-tested/3The benchmark uses strange settings. It does not even work properly: http://forums.techarp.com/reviews-articles/26957-x264-hd-benchmark-5-0-a-5.html#post369221

[Groucho2004]

The benchmark uses strange settings. It does not even work properly: http://forums.techarp.com/reviews-articles/26957-x264-hd-benchmark-5-0-a-5.html#post369221
The techarp guys have no clue (I already ranted about this in a different context (http://forum.doom9.org/showthread.php?t=165949):)). Yet, all the review sites (Anand, Xbitlabs) use the "benchmark".

[ChiDragon]

Anyone have some current recommendations for most FPS/$ spent? :)

Also interested in single-threaded performance.

[hajj_3]

Anyone have some current recommendations for most FPS/$ spent? :)

Also interested in single-threaded performance.

intel obviously wins single threaded. Check anandtech, they reviewed haswell and richland recently doing x264 tests.

[Juce]

Results with Haswell, Ivy Bridge and FX-8350: http://muropaketti.com/artikkelit/prosessorit/intel-core-i7-4770k-core-i5-4670k-haswell,4

x264 r2334. Settings:
video.y4m -o nul --quiet --crf 18 --tune film --preset medium
video.y4m -o nul --quiet --crf 18 --tune film --preset veryslow

[benwaggoner]

Results with Haswell, Ivy Bridge and FX-8350: http://muropaketti.com/artikkelit/prosessorit/intel-core-i7-4770k-core-i5-4670k-haswell,4

x264 r2334. Settings:
video.y4m -o nul --quiet --crf 18 --tune film --preset medium
video.y4m -o nul --quiet --crf 18 --tune film --preset veryslow
Looks like a decent improvement. Haswell is definitely the architecture to go for when it comes to $/GHz.

Of course, x264 is embarrassingly multithreaded; my 12-core Sandy Bridge can easily saturate all 24 virtual cores doing 1080p encoding. For most scenarios optimizing for GHz*cores*per-clock-efficiency is the best bet. I'm not sure what the exact numbers for the gain from hyperthreading, but x264 takes as good advantage of it as anything I've seen.

Something like this is probably great bang-for-the-buck today: http://www.amazon.com/Intel-i7-4770K-Quad-Core-Processor-BX80646I74770K/dp/B00CO8TBQ0/ref=sr_1_1?ie=UTF8&qid=1371226065&sr=8-1&keywords=haswell+core+i7.

I've not seen any >4 core Xeon options available yet. But for a multisocket system, this also seems like a fine deal: http://www.amazon.com/Intel-Chip-Processor-3-5-BX80646E31270V3/dp/B00D697QRM/ref=sr_1_1?s=pc&ie=UTF8&qid=1371226435&sr=1-1&keywords=E3-1270+v3

If I'm doing my math right, those should beat price/perf of any 2x6 core Xeon systems of the previous generation since the 6 core chips of similar clock speed are a lot more expensive

[mandarinka]

I've not seen any >4 core Xeon options available yet. But for a multisocket system, this also seems like a fine deal: http://www.amazon.com/Intel-Chip-Processor-3-5-BX80646E31270V3/dp/B00D697QRM/ref=sr_1_1?s=pc&ie=UTF8&qid=1371226435&sr=1-1&keywords=E3-1270+v3


That is an LGA 1150 (desktop-based) Xeon, and it obviously only supports single-socket configuration. Basically it is a Haswell sold as Xeon with premium (you got to buy a server/workstation motherboard too; the good part is that you can use ECC memory).

As for i7-4770K, I would instead buy i7-4770 or i7-4771 once it comes (100 MHz faster), unless you plan to overclock. The "K" thing is only good for OC, nothing more. Note that K chips also lack some features, like TSX (transactional memory) extensions and some virtualisation functions (vt-d).
The gain from HT is huge on Intel processors, but for HT you need to buy an i7 that is actually less good from the POV of price/performance ratio. So Core i5 is probably still a good or better choice despite lack of HT.

Of course, if Core i5 can't satisfy you, you will have to go for more expensive CPUs even if it is less cost-effective.

[benwaggoner]

Of course, if Core i5 can't satisfy you, you will have to go for more expensive CPUs even if it is less cost-effective.
And cost-effective isn't just about the CPU cost as well. Storage, display, CPU, etcetera can add up. I'm generally happy to pay 10% more for a total system in order to get 10% better total throughput, even if that means really spending 30% more on the CPUs themselves.

Of course, I'm the guy with 2x6 cores with 24 hyperthreaded cores already. Sometimes getting that one file done quickly matters a lot.