AMD Linux Gaming Performance Largely Unchanged With Linux 6.6 Git

[ptr1337]

skeevy420

Since the CO does lower your voltage also in idle scenarios, it could be that the CPU just gets to less voltage.
So, if the CPU does now less load on some CPUs due better handling of scheduling or whatever, it could result into a crash.
Actually often on the CO there is also a challenge having it in idle usage stable - my expierence with my Zen 3 if I had to low values.

For finding the correct CO values i can also suggest you "Ryzen Master". Yes, you would need to that time shortly windows, but after that the values are saved to your BIOS.
I really invested a lot of hours in my 5900X to find the correct values and then tested it with Ryzen Master and it was really near to the values which I got when doing all this stuff manually.
Two Cores had a bit higher value compared to my manual results, besides that it was equal.

But there are mixed reports about ryzen master. If you want to test it out, use the "per Core" Option and not the All Core one. The tests will run for around 1-1.5 hours.

[skeevy420]

skeevy420

Since the CO does lower your voltage also in idle scenarios, it could be that the CPU just gets to less voltage.
So, if the CPU does now less load on some CPUs due better handling of scheduling or whatever, it could result into a crash.
Actually often on the CO there is also a challenge having it in idle usage stable - my expierence with my Zen 3 if I had to low values.

That's the way it seems to me since a single click to select a desktop icon was enough to hard reset the system at -35 or more with EEVDF.

For finding the correct CO values i can also suggest you "Ryzen Master". Yes, you would need to that time shortly windows, but after that the values are saved to your BIOS.
I really invested a lot of hours in my 5900X to find the correct values and then tested it with Ryzen Master and it was really near to the values which I got when doing all this stuff manually.
Two Cores had a bit higher value compared to my manual results, besides that it was equal.

But there are mixed reports about ryzen master. If you want to test it out, use the "per Core" Option and not the All Core one. The tests will run for around 1-1.5 hours.

I've been meaning to do that but haven't since the goal of X3D overclocking is to prevent thermal throttling and I achieved that goal. I hit 87C stock and never above 85C with the lowered CO. Lowering the CO to -30 only dropped temps by 2C under full benchmark load and -40 didn't lower the temps further. My heat sync takes up so much room in my PC that I have to mount my side fan on the outside of my PC.

For my 7800X3D, a lot of the guides I read suggested starting out at -30 just to see what happens and to go one way or the other depending on stability. Since that's stable for me, I went up. I was about to go up to -45 with CFS when I tried EEVDF out and started getting crashes.

It seems like I have one or two cores that may not like anything over -35. Unfortunately, my MB doesn't have a Load Line Calibration setting (if it does, I can't find it) which can be set higher to offset voltage starved scenarios and help prevent crashing when CO is lowered too much. That's what happens when you buy a $120+tax MB for a $400 CPU. I paid $337, including tax, for my CPU in a New (Open Box) listing.

Yeah, my 7800X3D was cheaper than a 5800X3D. What's funny is that it was an angry impulse buy. After years of saving I finally, FINALLY, had enough money for a PS5 only to read that Sony was raising their online prices to $159 a year so I took my $800 in savings and voted with my wallet with an AM4 to AM5 upgrade. Fuck Sony and their ridiculously priced service.

[sdack]

This is a good test for the CCX-bound workqueues, but if the EEVDF makes any difference in gaming, I'd expect it to show up on on much lower core count CPUs. The 7950X is a 32-thread processor, and games almost never have that much concurrency. It makes sense that the scheduler would make effectively no difference if newly-runnable threads can always be punted to an idle CPU.

Agreed, but the results are still somewhat disappointing. Many of the benchmarks shown here will be more CPU- than GPU-limited judging by the high framerates these produce. I was expecting the improvements to the work queues with regards to L3 cache locality would trickle down to some extend to the L2 and L1 caches and show as smaller gains for a number of benchmarks. Now I am looking forward to a broader benchmark comparison and ideally with a couple more processors, i.e. the Ryzen 3, 5 and older Ryzen 7 perhaps.

[freerunner]

I'd like to know how the kernel has been configured wrt CONFIG_HZ. Is this info obtainable somehow?