If you found some issues here or have other optimizations which might be useful then you can create a Issue/PR if you want to.
Furthermore NOTHING of this is mendatory and you are free to apply and use what you want. I am just documenting what i found, use and some things which made problems for me.
This Repo has descriptions of how to optimize the gaming experience on Linux with AMD cpu and amd gpu.
This configuration i use is NOT meant to save power or be extra secure. Its also NOT meant to drastically improve performance.
The most imporant part i want to achieve is lower input/output latency and consistency !
Because a few more FPS won't matter at all, you might not even notice it, but when we optimize for consistency then we have the goal to keep the frametime as steady as possible and that is very noticeable.
Like the latency, the latency for the mouse or generally the system is also very noticable but not needed at all for server, so this is for desktop use only.
Even when you still can use some stuff fore the server, a lot of it would need to be changed, even into the other direction (so for more throughput instead of latency).
Latency and throughput are always tradeoffs. It either goes into one direction or the other, so you will have to decide for yourself.
For context i will show my pc specs here.
Operating System: CachyOS Linux
KDE Plasma Version: 6.6.4
KDE Frameworks Version: 6.25.0
Qt Version: 6.11.0
Kernel Version: 7.1.0-rc2-273-tkg-eevdf-llvm-ga293ec25d59d (64-bit)
Graphics Platform: Wayland
Processors: 16 × AMD Ryzen 7 5800X3D 8-Core Processor
Memory: 32 GiB of RAM (31,3 GiB usable)
Graphics Processor: AMD Radeon RX 7900 XTX
Manufacturer: Gigabyte Technology Co., Ltd.
Product Name: X570 AORUS MASTER
System Version: -CF
I also use the sched-ext sched PANDEMONIUM right now in adaptive mode. Which works like a charm.
Lets start off with bios settings. These play a really big role in performance.
First of all you want to enable your XMP profile for your RAM.
Then do these settings:
Resizeable Bar - Auto/Enable
Above 4G Decoding - Enable
Amd fTPM - Disable There has been some bugs where this caused some stutters, but i don't know if this is fixed already. Better to turn it off.
CSM Supoort - DISABLE !!! This is needed in disable for resizable bar support !
Core Performance Boost - auto/enable
CPPC - ENABLE
CPPC preferred cores - ENABLE
Global C state Control - ENABLE
AMD cool & quiet - ENABLE // this and some others are very important to being able to use the amd-pstate or amd-pstate-epp cpufreq driver
and right around amd cool & quite also needs to be something like this:
PState - PState0
Also dont forget to set the Cooler graphs of your fans in the case and cpu cooler in the cooler section in the bios.
This command should say more then 128MB
sudo dmesg | grep -i 'BAR.*VRAM\|amdgpu.*BAR'
output example:
[ 5.237537] amdgpu 0000:0d:00.0: [drm] Detected VRAM RAM=24560M, BAR=32768M
Here it is enabled.
you can easily see this in mission-center, or other task managers, but lets do this also in commandline.
inxi -mxx
and you will see something like this:
Memory:
System RAM: total: 32 GiB available: 31.25 GiB used: 11.76 GiB (37.6%)
Message: For most reliable report, use superuser + dmidecode.
Array-1: capacity: 128 GiB slots: 4 modules: 4 EC: None
max-module-size: 32 GiB note: est.
Device-1: Channel-A DIMM 0 type: DDR4 size: 8 GiB speed: 3600 MT/s
volts: 1 note: check manufacturer: G.Skill part-no: F4-3600C16-8GTZNC
Device-2: Channel-A DIMM 1 type: DDR4 size: 8 GiB speed: 3600 MT/s
volts: 1 note: check manufacturer: G.Skill part-no: F4-3600C16-8GTZNC
Device-3: Channel-B DIMM 0 type: DDR4 size: 8 GiB speed: 3600 MT/s
volts: 1 note: check manufacturer: G.Skill part-no: F4-3600C16-8GTZNC
Device-4: Channel-B DIMM 1 type: DDR4 size: 8 GiB speed: 3600 MT/s
volts: 1 note: check manufacturer: G.Skill part-no: F4-3600C16-8GTZNC
here you can then check if the desired ram speed is applied.
with this command you can see your used cpu freq driver. For amd you want to see amd-pstate or amd-pstate-epp
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver
output example:
amd-pstate
This will be very imporant for the kernel commandline parameter amd_pstate, which can be passive,active or guided.
Here the best performant is active. Which lets the hardware handle boosting of the cpu frequency BUT this does not always work, some hardware just cannot boost itself.
That is why its important to have some tools like mission-center with which you can check if the cpu frequency is boosted above the base clock. If amd_pstate=active is set and the cpu frequency is not being boosted above the base frequncy, then use amd_pstate=passive.
First of all i want to show my command line parameters:
amd-pstate=passive amdgpu.aspm=0 amdgpu.audio=0 nmi_watchdog=0 nowatchdog processor.max_cstate=1 transparent_hugepage=never vm.zone_reclaim_mode=0 audit=0 pcie_aspm=off ignore_rlimit_data split_lock_detect=off split_lock_mitigate=0 preempt=full libahci.ignore_sss=1 loglevel=3 rd.systemd.show_status=false transparent_hugepage_tmpfs=never amdgpu.dcdebugmask=0x4
amd-pstate as already described above tells the kernel if the kernel should use control the cpu frequency or the hardware itself. I have amd-pstate=passive because my hardware is not able to boost itself and is otherwise locked to the base frequency.
amdgpu.aspm=0 is used to disable the PCI Express power-saving for amdgpus otherwise the PCI Express link has a wake up latency when it goes into idle state and it wants to wakeup
amdgpu.audio=0 to disable the audio over HDMI/DP, i dont need it. I have an external DAC.
nmi_watchdog=0 is for Hard lockup detector and Soft lockup detection of Non-Maskable Interrupts, which uses a watchdog to detect that which needs to run in a set interval. When you disable it you free up some cpu time.
nowatchdog same as above but for all watchdogs
processor.max_cstate=1 sets the max C state or better said idle state to C1. It normally goes until C6 or even lower. The higher the number the deeper the sleep state. The deeper the sleep state the higher the wakup latency, to minimize the latency we do max cstate to 1 and i think you need the Global C state Control option enabled in the bios to really benefit from it.
transparent_hugepage=never disable transparent hugepages, normal pagesize is 4KB when THP is enabled every page is handles as a Hugepage and has around 2MB which might enhance the performance in some cases but it also leads to stutters in other cases and we want to optimize for smoothness so we disable it to reduce stutter/lag. When enabled the kernel starts a process in the background called khugepaged which tries to compress the THP pages in RAM, which causes stutters, lags and hangs and we dont want that.
vm.zone_reclaim_mode=0 tells the kernel to not reclaim already used ram from other application hastly
audit=0 disable logging of syscalls, logins, logouts, file access stuff and other stuff. TLDR: reduce cpu overhead
pcie_aspm=off same as amdgpu.aspm but more
ignore_rlimit_data tells the kernel to ignore resource limits for applications
split_lock_detect=off disables detection of splitlocks, which will be logged when detected with the other settings
split_lock_mitigate=0 when enabled the kernel throttels applications which makes splitlocks, when disabled the application will run at fullthrottle and the kernel imposes no penalty
preempt=full sets the preemption model for the kernel reduces latency of input/output devices But needs kernel which is compiled with the PREEMPT_DYNAMIC option enabled. But every kernel should have this by default so no worries. you can check with uname -a
libahci.ignore_sss=1 tells the SATA controller to ignore the Staggered Spin-Up bit for HDD and also has some benefits for SSDs TLDR: faster spinup and less latency.
loglevel=3 tells the kernel to only output logs with level 3 or higher at boot process
rd.systemd.show_status=false Hides service start messages during the ramdisk phase.
transparent_hugepage_tmpfs=never Disables THP tmpfs (temporary file systems stored in RAM) So that there are no files which are stored as THPs in RAM otherwise same as with transparent_hugepage happens.
amdgpu.dcdebugmask=0x4 disable Display Stream Compression DSC. For this you would need to check if the DP or HMDI cable you are using and the port on your Monitor and GPU are fast enough for the resolution and refresh rate your are using. Because it could be that you can use the current configuration of resolution and resfresh rate only with DSC. So in this case you should not disable it.
First of all i want to show you my sysctl:
vm.swappiness=1
net.core.busy_read=50
vm.max_map_count=2147483642
vm.vfs_cache_pressure=50
vm.dirty_ratio=80
vm.dirty_background_bytes = 67108864
net.ipv4.tcp_mtu_probing=1
vm.page_lock_unfairness=3
kernel.printk_devkmsg=off
vm.stat_interval=10
vm.zone_reclaim_mode=0
vm.compaction_proactiveness=0
vm.overcommit_memory=1
kernel.threads-max=1073741823
kernel.split_lock_mitigate=0
vm.dirty_writeback_centisecs=60
net.core.rmem_max=16777216
net.core.wmem_max=16777216
net.core.rmem_default=8388608
net.core.wmem_default=8388608
vm.unprivileged_userfaultfd=1
kernel.nmi_watchdog=0
kernel.unprivileged_userns_clone=1
kernel.printk = 3 3 2 3
kernel.kptr_restrict = 1
net.ipv4.tcp_congestion_control=bbr
net.core.netdev_max_backlog = 16384
Some of these like vm.dirty_ratio=80, net.core.rmem_default=8388608 and net.core.wmem_default=8388608 are more meant for systems with higher RAM amount like 32GB because they essentialy make programms use a little more ram in the case of the receive and sende buffer, or profit from more available/unused ram like vm.dirty_ratio=80 but you are free to adjust them.
The only thing i can recommend when you adjust vm.dirty_ratio then maybe also think about vm.dirty_background_bytes and vm.dirty_writeback_centisecs.
vm.swappiness Says how strong the pressure is to put stale stuff in memory into Swap the lower the less the pressure
net.core.busy_read=50 By setting this value to 50 (which represents 50 microseconds), you are telling the kernel: "When a process asks to read from a network socket and no data is there, don't put the process to sleep immediately. Instead, keep the CPU actively looping (polling) for up to 50 microseconds to see if data arrives."
vm.max_map_count=2147483642 extend max available Virtual Memory Areas per process
vm.vfs_cache_pressure=50 lower memory reclaim pressure from reclaiming memory used for caching directory and inode objects
vm.dirty_ratio=80 kernel can use up to 80% of the ram for delaying of writing of files to disk before hanging and writing everything to disk. The pro is smoother operation but the downside is possible dataloss on powerloss.
vm.dirty_background_bytes = 67108864 bytes from what point on the delayed file writes to disk from ram start in the background without hanging the system
net.ipv4.tcp_mtu_probing=1 tells pc to probe mtu size over the network
vm.page_lock_unfairness=3 this tells how often the same cpu core can steal the lock to a specific memory page one after another if that memory page is shared between more than 1 core. The lower the value the more fair the locking of the memory page is and less latency and the higher the more unfair and more throughput. Default is 5, i set it to 3 for nice tradeoff.
kernel.printk_devkmsg=off disable kernel's logging system for userspace applications
vm.stat_interval=10 This parameter controls how often (in seconds) the kernel's virtual memory (VM) subsystem wakes up to recount and update its internal statistics.
vm.zone_reclaim_mode=0 same as the command line parameter, maybe you dont even need this here if you have the commandline parameter
vm.compaction_proactiveness=0 disables ram memory page defragmentation, this disables kcompactd which repeatedly checks and defragments the ram. Which can cause stutter, lag and hangs.
vm.overcommit_memory=1 tell applications they can commit as much as they want. TLDR: Always say yes to memory allocations.
kernel.threads-max=1073741823 sets max system thread count.
kernel.split_lock_mitigate=0 same as in kernl parameters, very likely not even needed here
vm.dirty_writeback_centisecs=60 tries to flush dirty pages each 0.6 seconds to the disk. This is a lot more often than the default but this is the best option when combined with vm.dirty_ratio=80 where we never really are hanging hard to write everything but with vm.dirty_background_bytes = 67108864 which is very low when we start writing stuff to disk early and frequently but always in the background, so we keep stuff consistent and can guarantee no lag spikes, sutters or hangs from this. It does not happen until 80% of ram is full with it, which will never happen. Btw the kernel can reclaim these dirty memory pages all the time when ram is needed.
net.core.rmem_max=16777216 set max TCP/UDP receive buffer size.
net.core.wmem_max=16777216 set max TCP/UDP send buffer size.
net.core.rmem_default=8388608 set default TCP/UDP receive buffer size.
net.core.wmem_default=8388608 set default TCP/UDP send buffer size.
vm.unprivileged_userfaultfd=1 lets non-root applications to handle their own page faults, its is faster, but is also a security risk. So decide for yourself. But it brings less stutter and faster performance.
kernel.printk = 3 3 2 3 limits log messages to the important stuff, less overhead
kernel.kptr_restrict = 1 restricts viewing of kernel pointers to application which have CAP_SYSLOG capability or have root rights.
net.ipv4.tcp_congestion_control=bbr tcp congestion control algorithm bbr, better latency
net.core.netdev_max_backlog = 16384 increasing buffer where the kernel stores packets after they’ve been pulled off the physical network card (NIC) but before the CPU has had a chance to process them.
There are a lot of different schedulers which can help for specific usecases you might have. The default EEVDF is more build for throughput than responsiveness and gaming.
So for gaming i would recommend LAVD in performance mode or PANDEMONIUM, you can find them in the scx-scheds package in arch. Or just build them from source for example: https://github.com/sched-ext/scx or https://github.com/wllclngn/PANDEMONIUM
By default cachyOs uses scx-manager which is a gui which makes configuring them easier.
What you have to look out for is that you don't use ananicy while using most of the scx sched-ext scheduler because ananicy will change nice levels, io levels and so on. CachyOs has ananicy enabled by default. Here i would recommend to only use it when you use EEVDF scheduler and not with scx-scheds
You can disable ananicy with: systemctl disable --now ananicy.service
We also want to optimize the GPU performance.
With lact this is very easy.
just dowload it and then max the power usage limit to max or however you desire.
And for the most imporant part, you need to set the performance level to manual and then the power profile mode to 3D_FULLSCREEN or COMPUTE.
Or even better CUSTOM when you want to configure it yourself.
Important is here to set performance level to manual otherwise power profile mode does nothing.
You can of course also play around with the other stuff in lact.
performance level to high does NOT always boost frequencies to the highest clock rate.
Nowadays its very easy to use custom kernels, because cachyos already offers a wide variety of kernels with different cpu schedulers and optimizations applied.
Therefore you can just use one of them or just try a few of them and see what fits your use case.
If you still want to compile your own kernel because you want to change some things but don't want it to be to complicated then i can only recommend: linux-tkg https://github.com/Frogging-Family/linux-tkg
They make it really easy to configure basic stuff in the customization.cfg file. There is even a way to very simply apply your own patched if you want to.
For the driver you can just use the default mesa one if you want to, except when you want to have bugfixes and other stuff faster then you are also free to try mesa-git
One the other hand when you want to really optimize it then i can recommend: mesa-git from the frogging family same as the linux-tkg. https://github.com/Frogging-Family/mesa-git
They also make it here very simple to configure it and to compile it yourself.
Here you can apply cpu optimizations and other stuff which you might not have otherwise.
Another thing i noticed is that some application like firefox (in my case librewolf) set some threads/processes to Realtime priority, which can then not be handelt by EEVDF or sched ext schedulers and can lead to a lot of stutters.
And they use RTKit to set this RT priority. So we can just mask it with sudo systemctl mask rtkit-daemon.service but pay attention that your compositor still has RT prio, to have low latency otherwise that can be a problem. For KDE plasma this is no Problem because it sets its PRIO to RT with the Capabilites of linux. So they are not dependent on rtkit.
The only thing is that pipewire also uses rtkit and will now not have no RT prio. But this should not be problem with LAVD or PANDEMONIUM.
But if you still have issues then try increasing the Quantum size of the buffers in pipewire.