Electrically efficient CPUs/systems for x264/x265 encoding (but still fast)?

[Stereodude]

What are the most power efficient CPUs that are reasonably fast for x264/x265 encoding that can run Windows (using AVIsynth+ to feed x264/x265)?

I have a few year old R9-5950X build. It's pretty fast at encoding, but I was very disappointed when I recently measured its idle power consumption at ~100W. Under load it jumps to 250W+. Apparently I can lower the idle power consumption some by running the RAM at stock JEDEC frequencies (instead of the XMP timings) and making some other adjustments in the BIOS, but per reports I've read online it's never going to be great for a variety of reasons inherent to the platform.

On the other end of the spectrum, my HP Elitebook 840 G6 laptop with a i7-8665U is actually more power power efficient in terms of FPS of encoding per watt than the R9-5950x. It's also a lot slower. It uses 30W full at load only a few watts at idle, but is about 1/6th as fast at encoding as the R9-5950X. It makes it impractical for long high res jobs.

From my research online it seems the R9-7945HX in a SFF miniPC is probably one of the best options currently available in terms of efficiency in a high end (performance) CPU. From what I was able to gather online, it seems that it is both slightly faster than my R9-5950x at x264/x265 encoding while consuming about half the power under load. Idle power consumption is considerably better than ~100W at ~30W but that is somewhat dependent on what hardware the R9-7945HX CPU is in.

However, with AMD's CES 2025 announcement of the R9-9955HX, buying a R9-7945HX right now before seeing what the Fire Range R9-9955HX offers seems shortsighted. Intel doesn't seem to have a really competitive offering that I could find though maybe looking at high end "mobile" chips is not where efficiency has its peak. Maybe mid tier SKUs offer a better tradeoff of performance per watt while not being that much slower.

Are there some good CPU/HW options that I'm overlooking?

[Z2697]

"Retired" server parts?

[Stereodude]

Originally Posted by Z2697

"Retired" server parts?

My firsthand experience with those is that they tend to be rather power inefficient compared to modern consumer CPUs, but I'm talking about Xeon E5 V2 and V4 vintage hardware, not anything newer. What hardware do you have in mind?

[Emulgator]

Power efficiency can be achieved by choosing a Laptop with the tiniest production node CPU (any nm down),
no nVidia aboard, or deactivated (these are power greedy, I have one here),
and if tolerable encode on Intel iGPU from UHD 750 and up, plus simple underclocking.
While delivering comparable quality of a nVidia RTX3080, only 30.50% consumption.
I do this here for the quick-and still-goods.
If it's got to be CPU because of the last 10% quality, well then lets not talk about power efficiency anymore:
A CPU encode eats 800..1200% of an iGPU encode here, 14nm i9-11900K, and well,
this can be mended with any smaller node 10..7..5nm CPU
(Ryzen north of 5950 being good for that, Intel may still be greedier because of larger geometry)
but still: Encoding efficiency vs. Power consumption would still point to any 30W iGPU, besides of being magnitudes quicker.
R9-7945HX -> 9955HX: 5nm -> 4nm @ 55W, well, at this processing power this is very efficient anyway.
downclocking is still possible, base clock is 2,5GHz.

[rwill]

Originally Posted by Stereodude

What are the most power efficient CPUs that are reasonably fast for x264/x265 encoding that can run Windows (using AVIsynth+ to feed x264/x265)?

I have a few year old R9-5950X build. It's pretty fast at encoding, but I was very disappointed when I recently measured its idle power consumption at ~100W. Under load it jumps to 250W+.

[...]

Are there some good CPU/HW options that I'm overlooking?

Strange, a 5950X System should be ~50W idle max and ~175W under load...

Anyway 32 or 64 Core Threadrippers are currently the most energy efficient CPUs I guess, if you are able to utilize all Cores and do not overclock so it stays at around ~400W. But then again a 16 Core Ryzen Desktop at 175W is more or less the same, just a little less than half the compute and power...

[Stereodude]

Originally Posted by rwill

Strange, a 5950X System should be ~50W idle max and ~175W under load...

I'm sure the GPU in the system doesn't help the idle power, but the R9-5950X doesn't have an iGPU so it's somewhat unavoidable. From what I read on Reddit, the AM4 platform and the R9-5xxx CPUs disable much of their power saving features (like idling clocks and the like) if the RAM isn't run at the base 3200MHz DDR4 JEDEC timing (which hardly anyone does). I haven't bothered to spend the time to test all the BIOS settings yet.

The motherboard makers probably aren't designing all their multi-stage VRMs with power efficiency in mind either. So idle power draw from one AM4 system to another may vary greatly.

[Emulgator]

With the advent of few-mΩ MOSFETs these later VRMs can deliver in excess of 90%, even at today's Vcore below 0,8V.
Here an eight-phase system delivers 245A for the CPU's Vcore alone.
There is no extra VRM heatsink, the MOSFETs have just their PCB mount cooling.
This tells a good efficiency. The VRMs are not guilty, the user often is ;-)
Even departing from base timing will have to inhibit other saver algos, as any overclocking does,
going there would contradict energy efficiency, as full throttle will always do.
Losses rise exponentially, and notching up RAM speed 1 step will gain nothing but missing the signal's eye of validity more often in a while.
Those 33GHz scopes can tell...I went back from even so slight OCing after losing an expensive MB at a simple contradiction:
Choose silent mode and run SuperPi watching the maker's unfinished BIOS settings collide with semiconductor physics.
It happened within 15s. The implemented throttling on the MB was not sufficient to save it from harm.
As other encoders in RipBot subforum found out, limiting CPU to just below 100% made their systems just a few % slower, but much more silent, and less consuming.
Besides: At nodes below 14nm nominal I would rather not incur any thermal diffusion damage to the tiny junctions, rather stay way from too high current densities.
If everything cooks at +98°C, a hot spot may develop quicker than any fan/heatsink can convey.
Just +85°C and all gates survive.

[Stereodude]

I messed around with my R9-5950X system today and confirmed by turning off XMP, and unplugging the 3.5" HDDs that I was able to get the idle power consumption down to 50W from the wall. This is with a RTX 3070Ti in it. The four 7200RPM drives add about 25W on top of the 50W idle draw. Enabling the XMP for the RAM adds 20-25W to the idle power consumption taking it to about 100W.

[Emulgator]

Sounds resonable.

[Stereodude]

*sigh* Now the system is unstable and crashes. Now I have to figure out which power saving setting or settings in the BIOS are responsible.

[Z2697]

I mean more recent generations, for example EPYC 7002 or 7003, some of the models can be seen as counterparts of 32 or 64 core Threadripper that rwill mentioned, but typically not boosting as high (thus probably more efficient).

You can also try buying a bunch of mini PCs.
Talking about mini PC, maybe the new Mac mini can be a good choice too?

[Stereodude]

Originally Posted by Stereodude

*sigh* Now the system is unstable and crashes. Now I have to figure out which power saving setting or settings in the BIOS are responsible.

It turns out that even though I disabled PBO in the BIOS the negative curve offset, a PBO setting that I was playing with earlier in the day, was still active. This lowered Vcore from the default and made the system unstable.

It seems that enabling PBO and lowering the PPT wattage from the default PPT value is the best way to lower the power draw under load while still having dynamic clocks and not clamping single thread performance. It seems like quite a few watts can be saved under load with minimal performance loss. Of course with an idle power draw of ~75W slowing the system too much will be counteracted by the baseline power draw that's not accomplishing anything toward encoding.

[GeoffreyA]

Lowering PPT might also help.

Edit: Sent this before seeing your last post.

[Stereodude]

Originally Posted by Z2697

I mean more recent generations, for example EPYC 7002 or 7003, some of the models can be seen as counterparts of 32 or 64 core Threadripper that rwill mentioned, but typically not boosting as high (thus probably more efficient).

I haven't looked at EPYC. I'm not the biggest fan of trying to saturate a huge number of cores with parallel encodes or parallel split encodes, but if the efficiency is there...

You can also try buying a bunch of mini PCs.
Talking about mini PC, maybe the new Mac mini can be a good choice too?

I've seen reports that the M4 Mac minis are the most efficient hardware to run x264/x265 on. I don't think that my AVIsynth+ filters and flow are going to work on MacOS though.

[qyot27]

Originally Posted by Stereodude

I've seen reports that the M4 Mac minis are the most efficient hardware to run x264/x265 on. I don't think that my AVIsynth+ filters and flow are going to work on MacOS though.

Depends on which filters you need. Just because the list (points at signature) doesn't have macOS listed for a plugin doesn't mean it won't work, just that I haven't had the time to check it.

The same rationale might also kinda sorta apply to a cluster of Raspberry Pis, but at that point you're really going to hit diminishing returns due to all the other limitations of the SoCs.

[excellentswordfight]

Techpower up has an efficiency graph that is relevant here, its calculated on cinebench mind you, but the performance ranking does not differ that much from the x265 benchmark so its a decent substitute. The reason why the 7000-series looks more efficient is cause the 9000-series has a higher power-target, and usually efficiency goes down with more power. This can be countered by just enforcing a lower power target on the system. For examples I use an 12700k, that I have set long term boost to 125W, instead of 200W that is used for short term boost, I loose about 10% performance, for almost half the powerdraw.

So the best option for efficiency, if speed is still important, is 7000 and 9000 series Ryzen with powerlimits at max 100-125W (or a low TDP model), for 1080p and below 8C/16T is probably fine, and for UHD or multiple jobs (i.e. loads that can saturate the threads) efficiency increases for 12/24 and 16/32 models. However you are correct in your findings, desktop CPU:s and especially AMD has high idle power draw, for that either try to make power savings optimizations or just put it to sleep between jobs, or look at low-power-systems (but then throuput might be an iussue).

Originally Posted by Stereodude

I've seen reports that the M4 Mac minis are the most efficient hardware to run x264/x265 on. I don't think that my AVIsynth+ filters and flow are going to work on MacOS though.

Yoú could look at NUC/SFF system if you require x86, there should some models with AMD Ryzen AI 9 HX 370 out there, that would probably be your most efficient x86 option in this powerrange.

[Stereodude]

Thank for the informative post excellentswordfight.

For now I'm rolling with my 5950x with the TDP turned down a bit to 95W and with the HDDs unplugged (saving ~25W there). Performance loss was a few percentage from my quick tests, but power savings from the wall with the lowered TDP was about 40W+. Of course I blew it all up by running BM3DCUDA on the 3070Ti as part of the processing chain feeding the encoder so now the system is drawing even more power than before.