x265 on Core Ultra 285K?

[asarian]

ChatGPT told me I can expect a ca 20% slower performance on my P-Cores with the new Core Ultra 285K I am planning to buy.** (My old i9 12900k system is starting to faiL due to IMC-fatigue).

This performance drop is obviously no good. But it begs the overall question as to how to do x265 'properly' on the Core Ultra 285K. See, currently I let x265 render on the P-Cores only, as the parallelization across all cores (P and E alike) slows down the process even (as the entire process gets the performance of the lowest comnon denominator: the E cores).

I am surprised x265 has ACTUALLY never dealt with the P- and E Cores at all. x265, near as I can tell, can see multiple physical CPU's as different render-groups, as it were, but not between P and E Cores.

So, what is the best strategy for running x265 on Core Ultra 285K? Are there options I missed perhaps?

** The slow-down is reportedly due to the genius who decided to remove hyperthreading.

[rwill]

Wasn't the 12900k released 3.5 years ago? Is this now the expected lifetime of an Intel CPU thats not run at idle only? And you want to get another one from them AGAIN?

[asarian]

Wasn't the 12900k released 3.5 years ago? Is this now the expected lifetime of an Intel CPU thats not run at idle only? And you want to get another one from them AGAIN?

No, what I want is an AM5 platform, but, while there are very good Z890 motherboards, on AMD's side it's dismal: endless memory compatibility issues, boot loops, USB4 failures, etc.

[Z2697]

No, what I want is an AM5 platform, but, while there are very good Z890 motherboards, on AMD's side it's dismal: endless memory compatibility issues, boot loops, USB4 failures, etc.

Maybe you are reading it in an Intel circle.
I don't see how they can exist let alone (out)compete if they have such enormous problems.

If you want to hear it from me, I have use several generations of Ryzen CPUs and they do work fine and are very stable (just like any CPU should be, including Intel's).

[asarian]

Maybe you are reading it in an Intel circle.
I don't see how they can exist let alone (out)compete if they have such enormous problems.

If you want to hear it from me, I have use several generations of Ryzen CPUs and they do work fine and are very stable (just like any CPU should be, including Intel's).

My guess is because Internet reviews just look at the product-du-jour, and not at long-term stability, or user feedback since release.

Wanted to get an MSI X870E CARBON WIFI Nova, for instance, but even as short as 2 weeks ago, people are still reporting serious issues with it on reddit. The Intel Z890 platfoms offer better stability, less memory compatibility issue, but only have one drawback: they're made for a crappy CPU. :)

[asarian]

Maybe you are reading it in an Intel circle.
I don't see how they can exist let alone (out)compete if they have such enormous problems.

If you want to hear it from me, I have use several generations of Ryzen CPUs and they do work fine and are very stable (just like any CPU should be, including Intel's).

Oh, I believe you. CPUs from AMD are fine. :) It's the AM5 motherboards that are currentlly subpar, it seems.

[Z2697]

I never used a motherboard that high-end, is it more expensive than basically any CPU that fits in the socket?

But from the 3 mid tier (or some may even considered low-end) motherboards from MSI, ASRock and ASUS I have used, MSI has given me most trouble I think, and I've get rid of it.
Could be just that specific one I got.

MSI MAG B650M MORTAR WIFI
ASRock X670E PG Lightning (with 9950X)
ASUS TUF GAMING B650M-E (not to be confused with B650-E) (with 7950X)

I do have some problems with RAM and not able to pass 6000 MT/s, but I have bought cheap RAM so I can't be sure who's to blame here.
(I don't know what's wrong with me, I just like cheapest price that I'm still comfortable with more :o)

[GeoffreyA]

MSI has given me most trouble I think, and I've get rid of it.

On AM4 at least, I've had a solid experience with my B450 Tomahawk. When I move to AM5 or 6, I think I'll go with another Tomahawk. MSI is slack on the BIOS-update side though!

[microchip8]

I use 10+ years only MSI mobos with AMD CPUs and never had an issue with a single one of them. Once I bought an ASRock which was a disaster and burned out within a year.

As for encoding with x265, go AMD. No, seriously. Just do it. No need for a high-end mobo for that. The MSI Pro mobos are very good and stable (speaking from experience).

[asarian]

I use 10+ years only MSI mobos with AMD CPUs and never had an issue with a single one of them. Once I bought an ASRock which was a disaster and burned out within a year.

As for encoding with x265, go AMD. No, seriously. Just do it. No need for a high-end mobo for that. The MSI Pro mobos are very good and stable (speaking from experience).

P and E core scaling was my original question actually. I don't really see that working with the 285K. When you run an x265 process, 1 thread is placed per core (24 here). That's where the parallelization goes wrong, because if the entire process is divided over 24 cores, all cores end up waiting for the slowest E-Core. So up until now I just turned off the E-Cores completely. But now that the 285K has 16 of them, you would want to use them. But x265 doesn't really allow that: you can put different tasks on different NUMA nodes, but P and E Cores are, of course, not seen as two separate NUMA areas.

So, yeah, I may need to go to AMD.

[Z2697]

I think that's more of a operating system's problem

[asarian]

I think that's more of a operating system's problem

How so? I'm using Windows 11, btw.

Way I understand it, x265 parallelizes the threads over the cores (1 per core, aka 24 threads here). Can this be overriden then?

[Z2697]

On AM4 at least, I've had a solid experience with my B450 Tomahawk. When I move to AM5 or 6, I think I'll go with another Tomahawk. MSI is slack on the BIOS-update side though!

I've used AM4 boards from MSI (X570 GAMING EDGE iirc) and even prefered it over an ASUS TUF.

But once I upgrade to AM5, MSI, at least that one board I have, really gives me some trouble, especially with DDR5.
Memory Context Restore is a mess even with latest BIOS update.
To be fair it wasn't a big issue for me, because I don't turn off or restart my main system often, but once it became a "secondary" system... really annoying.

[Z2697]

How so? I'm using Windows 11, btw.

Way I understand it, x265 parallelizes the threads over the cores (1 per core, aka 24 threads here). Can this be overriden then?

I don't know, I only had a P/E core system for several days and returned it (guess why).
But I think miocrochip8 suggests Linux.

In a similar thread:
https://forum.doom9.org/showthread.php?p=2018564#post2018564

[GeoffreyA]

I've used AM4 boards from MSI (X570 GAMING EDGE iirc) and even prefered it over an ASUS TUF.

But once I upgrade to AM5, MSI, at least that one board I have, really gives me some trouble, especially with DDR5.
Memory Context Restore is a mess even with latest BIOS update.
To be fair it wasn't a big issue for me, because I don't turn off or restart my main system often, but once it became a "secondary" system... really annoying.

I've read about the Memory Context Restore issue on AM5, causing booting to stretch to minutes if power was completely turned off. Is this poor design on AMD's part, or only affecting certain motherboards?

[Z2697]

I've read about the Memory Context Restore issue on AM5, causing booting to stretch to minutes if power was completely turned off. Is this poor design on AMD's part, or only affecting certain motherboards?

It seems to not affecting some motherboards from what I can tell.
Not sure about the ratio of "good" and "bad" ones.

[Asmodian]

I do not like E-cores on desktop systems. They cause more trouble than they are worth. I usually kept then off with my 12900K too.

I would not get a 285K for x265, hybrid architectures are not great. Get a 9950X or a 265K (fewer e-cores going to waste) instead.

I've read about the Memory Context Restore issue on AM5, causing booting to stretch to minutes if power was completely turned off. Is this poor design on AMD's part, or only affecting certain motherboards?

I can use Memory Context Resource without any issues, running 96GB (2x48) of DDR5 at 6400 MT/s with tuned timings (Asus X870E-E). Training DDR5 with tight timings takes a couple of minutes, but with Memory Context Restore enabled it keeps the training through power cycles unless you change memory settings (even if unplugged), so boot times are only a few seconds normally.

Maybe certain motherboards still have issues, but I think as long as you get Hynix A or M-die memory and a current BIOS, AM5 is stable now. The hardware forums I frequent do not have people reporting issues anymore.

Something like an Asus B850-I is a great option for an AM5 motherboard.

[asarian]

I am currently looking at the ROG STRIX B650E-E GAMING WiFi; very basic, but does support 9950X, plus 3x M2 (without running into bifurcation issues), ans stll has a spare PCIE3 slot for my discrete sound card. :)

[excellentswordfight]

Looks faster to me?

https://tpucdn.com/review/intel-core-ultra-9-285k/images/encode-h265.png

I also have a alder lake CPU, and I do not experience the P/E core issue you are describing, using both P/E works fine in terms of x265 throughput (although windows scheduler is a bit weird so it has some other issues), and if I need to use my computer for other stuff i just minimize the process and run which make the encode to run only on E-cores.

What OS are you running?

[tormento]

If you want to stick to Intel, buy a used Xeon. Or new, if you want to sell a kidney.

Or wait until next tick iteration.

[GeoffreyA]

I can use Memory Context Resource without any issues, running 96GB (2x48) of DDR5 at 6400 MT/s with tuned timings (Asus X870E-E). Training DDR5 with tight timings takes a couple of minutes, but with Memory Context Restore enabled it keeps the training through power cycles unless you change memory settings (even if unplugged), so boot times are only a few seconds normally.

Maybe certain motherboards still have issues, but I think as long as you get Hynix A or M-die memory and a current BIOS, AM5 is stable now. The hardware forums I frequent do not have people reporting issues anymore.

Something like an Asus B850-I is a great option for an AM5 motherboard.

It seems to not affecting some motherboards from what I can tell.
Not sure about the ratio of "good" and "bad" ones.

Asmodian, Z2697, thanks for the comments on Memory Context Restore.

[hellgauss]

I also had some issues on win11 with P and E cores, using command line ffmpeg.

How do you run x265? Try to run in standard shell, not windows terminal. To do so, either uninstall terminal or run ffmpeg/x265 as admin (it should run in standard shell rather than terminal). Or you can try to rise x265 priority to above normal.

[asarian]

I also had some issues on win11 with P and E cores, using command line ffmpeg.

How do you run x265? Try to run in standard shell, not windows terminal. To do so, either uninstall terminal or run ffmpeg/x265 as admin (it should run in standard shell rather than terminal). Or you can try to rise x265 priority to above normal.

Pardon my daftness, but what'a 'normal shell'? I typically run x265 in a cmd command window.

[hellgauss]

Windows 11 has a different shell named "windows terminal", an "advanced" command shell which behaves differnetly from the standard cmd window. To run the standard cmd window you have to run cmd or the .bat file as administrator, or uninstall windows terminal to force opening in standard cmd.

Or try what happens if you run the x265 task in abovenormal priority. You can set priority in task manager.

See
https://www.elevenforum.com/t/change-default-terminal-application-in-windows-11.3410/
https://forum.doom9.org/showthread.php?t=186312

In last thread I find a way to manage E/P core usage using advanced power options, but I no longer used the method since my laptop CPU get too hot. Now I simply uninstalled windows terminal and use "balanced power option".

I do not know if it will solve your issue, but you can do a try.

[Z2697]

You don't have to uninstall it. There is an option to select default shell.

[tormento]

Windows 11 has a different shell named "windows terminal", an "advanced" command shell which behaves differnetly from the standard cmd window.
Windows Terminal is a container not a shell in its own.

Consider it a nice dress for command prompt or whatever real "shell" you use inside it.

Much better to set the affinity of the process for intended CPU cores.

[Z2697]

Windows Terminal is a container not a shell in its own.

Consider it a nice dress for command prompt or whatever real "shell" you use inside it.

Much better to set the affinity of the process for intended CPU cores.

Windows Terminal is a container, and it also includes a shell. OpenConsole.exe
Or not really a shell. It's basically the same as conhost.exe. I don't really know what they should be called, more accurately.

[asarian]

You don't have to uninstall it. There is an option to select default shell.

I generally just do a 'Open PowerShell window here', then type 'cmd.'

[asarian]

Much better to set the affinity of the process for intended CPU cores.

Which is what I do now. I use Process Lasso Pro to tie x265 to P-Cores only, currently (And for light frame server jobs, I set VSPipe to run on the E-Cores.)

I was trying to use --pmode and --pme as option, hoping x265 can itself better spread out over all its Cores (having removed affinity, of course), but x265 said those options were deprecated, and aborted the process.

[Z2697]

Or just aviod mixed cores processor to save the headache, unless there's some characteristic that you really want (power efficiency for laptops as an example).

This even includes mixed conventional + compact Zen cores processor, and 2 CCDs with only 1 "X3D" CCD CPU.
I heard that the "core parking" issue with 7/9950X3D is annoying (for gamers).

[Z2697]

Which is what I do now. I use Process Lasso Pro to tie x265 to P-Cores only, currently (And for light frame server jobs, I set VSPipe to run on the E-Cores.)

I was trying to use --pmode and --pme as option, hoping x265 can itself better spread out over all its Cores (having removed affinity, of course), but x265 said those options were deprecated, and aborted the process.

pme and pmode is deprecated in 4.1 or 4.0, can't remember.
They are not useful at all, I guess that's the reason why they are deprecated.
You can still try them out using older versions.

[asarian]

Or just aviod mixed cores processor to save the headache, unless there's some characteristic that you really want (power efficiency for laptops as an example).

And this precisely why I started this thread saying "See, currently I let x265 render on the P-Cores only, as the parallelization across all cores (P and E alike) slows down the process even (as the entire process gets the performance of the lowest comnon denominator: the E cores)."

But with 16 E-Cores on the Core Ultra 258K (instead of 8 on my i9 12900k), plus hyperthreading gone on the P-Cores now, the penalty for ignoring those E-Cores becomes too large. Hence I was looking to see whether people have found smarter ways to spread things out over those P- and E-Cores.

[benwaggoner]

And this precisely why I started this thread saying "See, currently I let x265 render on the P-Cores only, as the parallelization across all cores (P and E alike) slows down the process even (as the entire process gets the performance of the lowest comnon denominator: the E cores)."

But with 16 E-Cores on the Core Ultra 258K (instead of 8 on my i9 12900k), plus hyperthreading gone on the P-Cores now, the penalty for ignoring those E-Cores becomes too large. Hence I was looking to see whether people have found smarter ways to spread things out over those P- and E-Cores.
I've found x265 on M-series Macs utilize the mix of core performance types quite effectively.

I'm not sure if there is logic to aim different kinds of threads at difference cores, but I imagine that would be effective on any architecture. Some stuff really benefits from powerful SIMD, some doesn't.

[Sagittaire]

Well It's always difficult to make comparisons.

Anyway in real world perfomance two Ecore (2C/2T) are at same performance than one Pcore with hyperthreating (1C/2T) but for really less electrique power (more efficiency). This is the main interest of Ecore.

But in monocore test, one Pcore is really more performant than one Ecore.

So Ecore will be better for video encoding if and only if you manage to saturate all the cores of your CPU to 100%.

And that's the problem: it's very difficult to saturate a CPU with many cores (16C/32T for example) with video encoding. And in this case, 8 Pcores will be faster than 16 Ecores.

[Sagittaire]

Anyway 285K will be at ~ same performance level than 9950X for video encoding. The 14900K is already only less than 10% below the 9950X.

When you have your 285K, make the test with my bench:
https://forum.doom9.org/showthread.php?t=185855