Hardware for dedicated x264 machine

[Lyris]

So, I currently have a machine which encodes video, almost 24/7. The CPU is an Intel i7 3770k, overclocked to about 4.5ghz from the stock 3.5ghz (there is a big Noctua NH-D14 cooler strapped on helping this happen). My understanding is that RAM is not that important here, so I only have 4gb in there. And of course, no need for a big GPU, I'm just using on-board video.

Depending on the amount of high frequency content in the video, I'm getting about 3.5fps on the "veryslow" mode. (Keep in mind this is going from uncompressed or nearly uncompressed masters, not transcoding already encoded video). A totally noise/grain-less video camera interview can easily sail through much faster.

Does anyone have experience using x264 with higher-end processors? Intel Xeons for example? If so, did you see significant improvements in encoding speed compared to "consumer" CPUs? Or should I just get whatever has the highest clock speed and to hell with anything else?

[mandarinka]

Combination of clockspeed and core count. Haswell-E might be a good upgrade later this year, but sadly they are only preparing hexacores in the cheaper parts, the octacore (base clock 3 GHz) is going to cost 1000 USD (plus not exactly cheap motherboard and DDR4 memory). So besides that (and current Sany Bridge-E/Ivy Bridge-E hexacores), you are looking at using Xeons...

x264 does scale beyond 4core/8thread CPUs, eight cores and probably 16 cores should still give pretty good scaling. (and of course, there is the option of running more encodes concurrently)

[benwaggoner]

Originally Posted by Lyris

Does anyone have experience using x264 with higher-end processors? Intel Xeons for example? If so, did you see significant improvements in encoding speed compared to "consumer" CPUs? Or should I just get whatever has the highest clock speed and to hell with anything else?

I've done quite a bit of encoding on 12-core Sandy Bridge and 16-core Ivy Bridge. I hit 100% at least half the time doing 1080p with --preset veryslow as long as the source decode isn't too bad. 16-core is a little harder to saturate, but I've done it with challenging source coming from a trivial-to-decode source (like Y4M from a SSD RAID 0).

Per clock per core a Haswell certainly outperforms Ivy Bridge, but there's a much lower limit on how many cores you can get. For harder to parallelize content, fewer faster cores is going to give better overall throughput.

[Atak_Snajpera]

Why do you waste space using uncompressed video files? Why don't you use ut video or magicyuv?

Whenever i have problem with core saturation i just encode multiple chunks at the same time.

[benwaggoner]

Originally Posted by Atak_Snajpera

Why do you waste space using uncompressed video files? Why don't you use ut video or magicyuv?

Whenever i have problem with core saturation i just encode multiple chunks at the same time.

I was also testing the same content with x265, which only has native support for .yuv and.y4m

[Atak_Snajpera]

Both we know that this is not a problem with avs2yuv tool.

[burfadel]

Originally Posted by benwaggoner

Per clock per core a Haswell certainly outperforms Ivy Bridge, but there's a much lower limit on how many cores you can get. For harder to parallelize content, fewer faster cores is going to give better overall throughput.

I saw a review earlier today of the i7-4790K (Haswell Refresh), it does appear to be faster than the older Haswell's. This is probably largely due to it being clocked higher. They supposedly will overclock higher as well, which will add to the benefit. It wouldn't be worth upgrading to it though if you already have a decent CPU. I think a good water cooler would go well with the K series Haswell Refresh.

Overclocking to 4.5 GHz is quite impressive on a i7-3770K, usually if you push them hard at that speed they'll overheat, regardless of how good the attached cooler is. This is because the cores simply can't transfer the heat quick enough to the CPU casing due to Intel cheaping out with the thermal interface material.

The only way people usually get to 4.5 GHz without overheating the i7-3770K etc when pushed is by delidding, but that is a potentially CPU-destroying exercise.

[Asmodian]

i7 3770K, 4.5 GHz, max temp seen ~68°C (Prime95 max heat).

But an i7-5960X finally looks like a good upgrade for a not too crazy expensive x264 box.

[Blue_MiSfit]

I administer a large number (100+) of high performance 12-16 core servers running a lot of different video encoding workloads, including x264.

Generally speaking, if you can feed x264 fast enough (i.e. if you're using no AviSynth or a light AviSynth script, and have a fast decoder), it will scale nicely up to use 12-16 cores for HD content on one of the slower presets.

CBR encoding used to be problematic with tons of threads (36 threads for a 12 core system with HT) and buffer sizes less than ~ 1 second, so I had to limit thread count in that case. I THINK this was fixed, but I'm not sure

[Groucho2004]

This thread is a good read on the subject.