Some AI model use only 1 CCX on Ryzen CPU with 2 CCX ( 5900X)

Hello everyone :slight_smile:
Windows 10
Ryzen 5900X
RTX 3070 Ti
64 Go RAM G.Skill F4 3600 C16
MSI b550 Tomahawk

Bug or problem, I can’t really define it.
I notice that some models (I see on Gaia and Nyx) use only 1 CCX (core 0>5)
It’s an issue for confort, temp regulation use global temp, only use 1 CCX up global temps for nothing.
It’s a performance issue when i do other task when Topaz video AI is working, the other task who use all CPU core are limited because Topaz Video AI saturated half of my first 6 core.

It’s a bug or a lack of optimisation, but it’s an unfortunate issue.

Screen of utilisation on only CCX 0 when Topaz running Gaia model

Hope all model can run on all CPU core/CCX in the futur.
Have a nice day !

In our testing we’ve found that spreading to about half of the CPU cores is the optimal setting for most systems. This prevents over-utilization and system unresponsiveness during processing.

Hello and thank you for your reply :slight_smile:
I totally understand what you are saying, the maximum load is well distributed between 3 cores.

So I’m attaching another, more explanatory screen.

Your optimisation is impeccable for single chiplet CPUs.
At the moment my 5900 is being used as a 3600 and personally I use my computer while Topaz Video is running.
When I’m using rendering software, especially 3D, it’s catastrophic in terms of both performance and temperature, where on average I’m 7 degrees warmer because of this truncated CPU usage.
As far as performance is concerned, I simply freeze at every action, which I didn’t have before.
On my side I can’t do any better, it’s already UV at -150Mv with a lower maximum power consumption, I can’t do any more without dropping performance enormously or even losing stability.

As far as possible, I’d like a choice equivalent to the GPU choice we have:

proc_proc

That I could have a “non-recommended” CPU usage choice that uses my processor as a 2 chiplet processor and not 1.

I know that the Ryzen 2 chiplet architecture is no joy to optimise, I’m not dreaming of something perfect, but it would be very much appreciated if a minimum effort could be made to load the CCX 1 a little more (as opposed to very little at present).
The V2 was 2/3-1/3 use CPU CCX and it was so better compare that V4 :roll_eyes:

Have a nice day :slight_smile:

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.