2990WX Threadripper Performance Regression FIXED on Windows* #threadripper | Level One Techs

That is why I would pitch Blizzard an email first asking if it was okay to run the programm (linking to the github) while doing tests (in the included trial mode).
Maybe they say “yes”, maybe they say “yes, but only on the trial server”, maybe they say “no”. I just find this curious and want to test if it helps (currently there are wierd FPS drops whenever the game does Disk writes).

Did Microsoft or AMD ever provide a response if this will be fixed in Windows without relying on this 3rd party fix?

If Microsoft were to not fix this, their presence in high performance servers would take a hit. What worth is a server that wastes half its power in context switches?

1 Like

As bad as Microsoft is at providing services its customers need, I would have to agree with MazeFrame, Microsoft will fix this problem. Unless fixing Windows performance regression requires a complete reworking of Windows kernel and Microsoft decides it isn’t worth fixing.

1 Like

Knowing Microsoft i don´t really see this being fixed any time soon.

They might threadripper could end up being a significant part of the enterprise environment for higher end workstations since not everything will use epyc realistically. I do not think Microsoft likes performance regressions… they usually try to fix them

1 Like

@wendell - out of interest did you get a chance to give ThreadWrangler a spin? Still interested to see what it does on 2990WX

Additionally, rumour has it that Microsoft are purchasing a lot of EPYC boxes for Azure, so i suspect they would really like this fix for themselves…

1 Like

With Intel having shot themself in both feet, Dell and HPE recommending their customers to buy Epyc systems instead, I can see why Microsoft would buy AMD systems.

Would be quite the embarrassment to admit “our OS performs like shit on our compute service”.

2 Likes

If nothing else, fixing it will save them half the number of boxes needed to handle a compute load of X.

1 Like