AMD Ryzen 7 1800X vs. Intel Core i7 7700K Linux Gaming Performance

[nuetzel]

If you want to try, you can boot with "pcie_aspm.policy=performance" as kernel parameter, this should give you the same result.

You can check with:

cat /sys/module/pcie_aspm/parameters/policy

It should now say "[default] performance powersave", with the parameter it should say "default [performance] powersave".
If there's no performance difference, than you're not running into this problem

But _only_ if your (distro) kernel support, that.
E.g. offer the 'missing' modules. --- openSUSE Kernel:stable (Tumbleweed) do NOT.

/opt/mesa> grep ASPM /boot/config-4.10.1-5.gf764d42-default
CONFIG_PCIEASPM=y
# CONFIG_PCIEASPM_DEBUG is not set
CONFIG_PCIEASPM_DEFAULT=y
# CONFIG_PCIEASPM_POWERSAVE is not set
# CONFIG_PCIEASPM_PERFORMANCE is not set

Some numbers would be nice!

Greetings,
Dieter

[hoohoo]

That's expected. High resolutions moves the bottleneck to the GPU rather than the CPU.

For the video cards - Fury vs GTX yes that makes sense. Why should it be the case when the independent variable is the video card and the dependent variable is CPU?

[dimko]

When I can see this clearly with my X4 845, with Ryzen you should see more clearly. Linux performance governor is hindering the speed of my custom kernel, even when the performance governor is used.

What exactly do you see, please elaborate. Did you manage to squeeze some performance from your machine by not using governor? I thought it's there by default and every Linux machine MUST use it?

[indepe]

If you want to try, you can boot with "pcie_aspm.policy=performance" as kernel parameter, this should give you the same result.

You can check with:

cat /sys/module/pcie_aspm/parameters/policy

It should now say "[default] performance powersave", with the parameter it should say "default [performance] powersave".
If there's no performance difference, than you're not running into this problem

Thanks, sounds good, will do.

[Tomin]

I am thinking amd64 based SBCs with pluggable ram, pluggable processors, m.2 bus for SSDs, pluggable wifi/NIC cards. (Maybe I have gone too far )

I don't think that is a SBC anymore.

Erm, should I say an SBC (es bii see or whatever you'd use) or a SBC (single board computer)?

[droste]

Is it possible to translate this into english? Do I need to worry about running into the same problem? (Once I have a system able to run Vulkan without recompiling the kernel, that is, which will be soon.)

If you want to try, you can boot with "pcie_aspm.policy=performance" as kernel parameter, this should give you the same result.

You can check with:

cat /sys/module/pcie_aspm/parameters/policy

It should now say "[default] performance powersave", with the parameter it should say "default [performance] powersave".
If there's no performance difference, than you're not running into this problem

[indepe]

You can not tune windows desktop, but Linux desktops you can. Use a non debug 300Hz kernel, disable cpu freq governor etc. If Phoronix needs games with threads, benchmark games with wine-staging, that will distribute load to cpu cores.

What kind of performance improvement are you measuring when doing this? Or is your suggestion specifically for Ryzen, and you don't have one (yet) ?

[zboson]

Apparently AMD has suggested to disable SMT for gaming benchmarks - seems to have made a marked difference for some games on the Windows benchmarks I have seen. It would also be cool to see how it compares to the FX-8370 like the previous article

http://www.anandtech.com/show/11170/the-amd-zen-and-ryzen-7-review-a-deep-dive-on-1800x-1700x-and-1700/10

"However, when running in SMT mode but only with a single thread, the statically partitioned parts of the core can end up as a bottleneck, as they are idle half the time."

That's an interesting point. There are cases when SMT in not helpful. For example in large dense matrix multiplication. If can take full advantage of a core with a single thread and using another thread will be less efficient. At least that is true with Hyper-Threading. So there are cases where you only want one thread per core. I wonder then if Zen with SMT enable is less efficient with one thread per core then it is with SMT disabled.

It's even more complicated on Knights Landing because it supports four threads per core. Some suggest to turn off Hyper-Threading on Knights Landing. I don't have enough experience yet with it know what is best. It of course depends on the application. Here is an interesting comment about four threads per core by the "father" of Xeon Phi

http://www.agner.org/optimize/blog/read.php?i=761#763

[agaman]

This maybe your opensource radv vulkan performance fix: Solution--reconfigure kernel: "make menuconfig"/Bus options/PCI support/PCI Express Port Bus support/PCI Express ASPM control/Default ASPM policy (set to "Performance")

Why only radv is afected by this? Can you see any improvement in OpenGL after this change?
It's radeonsi doing something different to change the PCI performance mode that radv isn't doing?

[sarfarazahmad]

this is somewhat offtopic. what would AMD ZEN microarchitecture based single board computers looks like ? Also what is single board computers be modular something like project Ara from Google. does this even make any sense ?

I am thinking amd64 based SBCs with pluggable ram, pluggable processors, m.2 bus for SSDs, pluggable wifi/NIC cards. (Maybe I have gone too far )

Also imagine a 16-core amd64 SBC, what use could that possibly have ?