Showing results for 
Search instead for 
Did you mean: 

Drivers & Software

Adept I

Wrong core affinity with the new DX11 driver and Zen 3 with two chiplets

I reported this problem already several times in the report tool of the AMD drivers with no avail, so I will try here to see also if more people can verify this.
My system: R9 5900X/RX 6800/MSI B550 Gaming Edge/48GB ram 3800Mhz
Well, the problem is that since the release of the May Preview driver (and still working like that with the 22.7.1), the main game threads of DX11 games go to the slow CCD or chiplet of my Ryzen, especially, those with low thread count. For example, I tested that in Assetto Corsa (720p minimum details to force cpu bottleneck and observe this better), my fast CCD is the 1, the slow one is CCD 2:
as you can see, all the loads are in the wrong CCD (before those drivers, the load was completely on the CCD 1, which allows 200-300Mhz more frequency). This gives 665fps:
I repeated the test after disabling CCD 2 in bios, obtaining 738fps:
which is what I should get without doing anything if the core affinity worked as intended. Of course, CPPC is enabled in bios. I hope that this gets some visibility and gets fixed somehow.
5 Replies
Adept I

That's a HUGE negative impact when using a two ccd chip on newest DX11 optimized drivers. I think this should be pretty easy to fix, hope some AMD responsible can take a look into this as soon as possible.


IMPORTANT NOTE using ryzenmaster and checking your CCX and individual cores some cores are golden sample sorta 'go a bit faster than other cores maybe'. so it may be referring to the optimal core instead of the first one you see or might expect. But the other reason may be the app might be 'running as a service' or in background or isnt maximized fullscreen exclusive while youre MONITORING SOFTWARE IS IN THE BACKGROUND.. you see idle tasks switch between cores and performance usage based on foreground and background to multitask and manage shared resources.

all dx9 dx11 arent multicore or multithreaded.. BUT they ARE single thread.

So in 60's if your giant slab of computer runs at its one setting all day every day it overheats.

So they made single thread software to be like 2% or 5% of whatever it SHOULD be running at so it doesnt ever get close to overheating.

They took a big fat pipe of bandwidth and CPU engine power and made it use nylon rope hairs and only ever a single one..

since most computers on the market at the time were fake ones and not computers at all as they had high latency and were designed to be affordable for warehouses databases and offices they were making them for inventory and cash registers and banks.. which being a government service and a required utility service was cheap as possible and had no money involved in the banking industry at all as it never ever turns a profit and runs at cost price bank accounts exist by LAW and you must have and use one by law as a crime prevention measure. so if the government and law force you to open a bank account then take the money that they're supposed to be keeping safe from criminals they go to prison. law says you cant have more than a few thousand dollars cash items or money in your house or on your person as it encourages criminals. So they need computers to track and manage it so they're all the cheapest lowest cost solutions that get the job done usually. many banks almost never upgrade their older custom hardware. So these cash register type computers that can still run commercial commerce software such as IBMs COBOL commerce code later scribbled all over and renamed to C and C++ (more of the same hair thin single thread of nylon woven rope) means that your CPU usage will SHOW extremely high on a single core and SAY that its using 70% .. but you will have millions and millions of times WORSE performance than when DX12 shows 8% on a single core.

a better question is how many of my radeon 5700xt's 40 compute cores are being used by the GPU in a game when it shows 10 to 20% in thousands of times beyond ultra quality octillions of resolution 3 kinds of ray methods mortal kombat aftermath 11 ultimate

You will want to have it run on ALL cores not just the one. try forcing to use direct3d renderer in directx12 mode or force the use of VULKANRENDERER. by launching it with arguments on the commandline maybe in a shortcut or in the steam launcher link or by creating a config.ini text file and specifying what renderer to use CAPABILITYVULKAN1000 VXR VMA and so on? enable your infinitycache and infinityfabric in there.. uhh specify simultaneousinstantaeousmultiplepasses i guess? enable exclusivemode? FS2? VRR ? augmentedenhancedultramaxfullspeed maybe?

enable maybe a few thousands of tessellation? or 256xtessellation or infinitytessellation.. use infinity resolution maybe?

uhh try using infinityEQsupersampling? 

disable ryzenmaster in task scheduler perhaps? then try again. 

uninstall all .net and windows desktop runtime and C++ from installed APPs and add remove components and go download the latest from microsoft website. 





Man, this is just that the threads are not loaded on the good chiplet with the cores with the highest boost as any other software that uses a couple of strong threads. That is all.

And this happens since the introduction of the new DX11 driver (May Preview, 22.5.2, 22.6.1, and 22.7.1) in DX11 games (another one would be Fall Guys, for example). Rolling back to the drivers prior to those, the loaded cores are the ones with the highest boost (thanks to the CPPC information). It is not overheating or anything like that.

There is some change in the scheduling policy that makes people with two-chiplet Ryzen performing worse than it should (in my case almost 10% lower), just because the load goes to the slow cores in the slow CCD/chiplet.

On Intel terms, it would be like loading the game threads on the E-Cores, while letting the P-Cores unused.

I do not use Ryzen Master BTW.

Adept I

New driver (22.8.1) tested, the bug is still there

Adept I

22.8.2 driver tested. Unsurprisingly, nothing was fixed.