CachyOS Performance Optimization Guide

For: Intel 8th–10th gen Core i5/i7, NVIDIA Turing/Ampere, 16GB RAM, CachyOS with Wayland. Kernel: linux-cachyos, PREEMPT, 1000Hz, EEVDF (default) + sched_ext.

A practical guide for getting the most out of mid-range gaming hardware on CachyOS. Every section explains what CachyOS already provides, what you should add on top, the exact configuration files to create, and the tradeoffs involved. Copy-paste-ready configs at the end.

1. What CachyOS Already Does
1. Before You Begin
1. Boot Parameters
1. CPU Tuning
1. sched_ext: Swappable Schedulers
1. Memory & Swap
1. Storage & Filesystems
1. GPU: NVIDIA
1. GPU Overclocking via LACT
1. Gaming Pipeline
1. Audio: PipeWire Low Latency
1. Network
1. System Limits & Sysctl
1. Services to Disable
1. Desktop & Wayland
1. Minimum Viable Tuning (The 80/20)
1. Verification Checklist
1. System Recovery (If You Break It)
1. Configuration File Reference

1. What CachyOS Already Does

CachyOS ships with aggressive performance defaults. Understanding these prevents you from double‑configuring, creating conflicts, or breaking things that are already optimized.

Kernel

The linux-cachyos kernel includes:

PREEMPT (full preemption), lower scheduling latency
1000 Hz tick rate, finer‑grained scheduler decisions
EEVDF, the default scheduler, not BORE (see the correction below)
CONFIG_SCHED_CLASS_EXT=y, sched_ext support, lets you load a different BPF scheduler at runtime
Clang ThinLTO, link‑time optimization during kernel compilation
x86‑64‑v3/v4 microarchitecture targeting
NTSYNC, kernel‑level synchronization for Wine/Proton (replaces fsync/esync)

If you install CachyOS with the default kernel, you already have all of this.

Scheduler correction: linux-cachyos defaults to EEVDF (Earliest Eligible Virtual Deadline First, mainline's scheduler since Linux 6.6), not BORE. BORE (Burst-Oriented Response Enhancer, a fork of EEVDF tuned for burst and interactive responsiveness) ships as a separate package, linux-cachyos-bore. Install it and reboot if you want BORE as your always-on scheduler. What both variants share is CONFIG_SCHED_CLASS_EXT=y, the sched_ext framework, which is a bigger lever than picking BORE vs EEVDF at install time since it lets you swap the whole scheduler at runtime with no reboot. See section 5.

Sysctl defaults

/usr/lib/sysctl.d/70-cachyos-settings.conf:

vm.swappiness = 100
vm.vfs_cache_pressure = 50
vm.dirty_bytes = 268435456
vm.dirty_background_bytes = 67108864
vm.dirty_writeback_centisecs = 1500
vm.page-cluster = 0
kernel.nmi_watchdog = 0
kernel.split_lock_mitigate = 0
kernel.printk = 3 3 3 3
kernel.kptr_restrict = 2
net.core.netdev_max_backlog = 4096
fs.file-max = 2097152

Systemd additionally sets:

fs.inotify.max_user_instances = 1024
fs.inotify.max_user_watches = 524288
vm.max_map_count = 1048576
net.ipv4.tcp_keepalive_time = 120

ZRAM

Compression: zstd
Size: equal to physical RAM
Swap priority: 100
A udev rule sets vm.swappiness = 150 when ZRAM activates

Transparent Hugepages

CachyOS enables THP aggressively:

enabled: always
defrag: defer+madvise
khugepaged/max_ptes_none: 409

This benefits tcmalloc‑based applications (Chromium, Proton games) while the max_ptes_none tuning prevents the memory bloat historically associated with always.

IO Schedulers

HDD → bfq
SATA SSD → mq-deadline
NVMe SSD → kyber

NVIDIA Module Options

options nvidia NVreg_UsePageAttributeTable=1
    NVreg_InitializeSystemMemoryAllocations=0
    NVreg_DynamicPowerManagement=0x02

PAT=1, faster CPU↔GPU memory transfers
InitMemAlloc=0, skips zeroing system memory before GPU use (small performance gain)
DynamicPowerManagement=0x02, runtime D3 for Turing mobile GPUs (irrelevant for desktops, harmless)

Systemd Tuning

Setting	Value
Journal max size	50 MB
Service start timeout	15 s
Service stop timeout	10 s
DefaultLimitNOFILE (system)	2048:2097152
DefaultLimitNOFILE (user)	1024:1048576

Other Built‑in Optimizations

Audio: rtkit daemon, audio group realtime priority (rtprio 99), CPU DMA latency device access, snd_hda_intel power saving disabled on AC power
SATA: ALPM forced to max_performance
HDD: hdparm -B 254 -S 0, no APM spin‑down
Watchdog modules: iTCO_wdt and sp5100_tco blacklisted
NTP: Cloudflare + Google time servers
DNS: systemd‑resolved

Bottom line: You start from a strong baseline. The remaining sections are what you add on top.

2. Before You Begin

Rule #1: Don't fuck up your system without a net. CachyOS defaults to BTRFS for the root partition. Use it.

# Create a snapshot before you touch anything
sudo btrfs subvolume snapshot / /@pre-optimization

# To roll back later (if you can't boot):
# Boot from USB, mount your drive, rename snapshots.
# Or just use the CachyOS bootloader menu's "Read-Only Snapshot" option if you have it.

Rule #2: Test one change at a time. Rebooting after applying the whole list is a recipe for not knowing what broke.

3. Boot Parameters

CachyOS uses Limine as its bootloader. Parameters go in /etc/default/limine and get applied with sudo limine-mkinitcpio.

If you use a different bootloader (GRUB, systemd‑boot), the parameters are the same, only the configuration mechanism differs.

Recommended command line

quiet mitigations=off nowatchdog nmi_watchdog=0 nvidia_drm.modeset=1
tsc=reliable clocksource=tsc intel_pstate=active preempt=full
split_lock_detect=off pcie_aspm=performance intel_idle.max_cstate=1
transparent_hugepage=madvise splash rw

Parameter reference

Parameter	What it does	When to skip
`quiet`	Suppress kernel log output at boot	Never, cosmetic only
`mitigations=off`	SEVERE, disable all CPU vulnerability mitigations	Skip if this machine handles sensitive data or runs untrusted code
`nowatchdog`	Disable all watchdog timers	Never, watchdog timers are unnecessary on consumer desktops
`nmi_watchdog=0`	Disable NMI watchdog specifically	Never, redundant with `nowatchdog`
`nvidia_drm.modeset=1`	DRM modesetting for NVIDIA	Required for Wayland. Only skip if using X11
`tsc=reliable clocksource=tsc`	Force TSC as clocksource	Skip on unstable TSC hardware (rare on Intel Core)
`intel_pstate=active`	Intel P‑State in active mode	Skip on AMD CPUs (not applicable)
`preempt=full`	Full kernel preemption	CachyOS kernel already has this compiled in, but adding it explicitly is harmless
`split_lock_detect=off`	Disable split lock detection	Skip on server workloads (split lock detection catches bugs, not relevant for gaming)
`pcie_aspm=performance`	Disable PCIe Active State Power Management	Skip on laptops (increases battery drain)
`intel_idle.max_cstate=1`	MODERATE, limit CPU to C1 idle	Skip on laptops or if power consumption matters
`transparent_hugepage=madvise`	THP on madvise hints only	Note: CachyOS tmpfiles overrides to `always` at runtime
`splash`	Plymouth boot animation	Never, cosmetic only
`rw`	Mount root read‑write	Always required

Security Warning: mitigations=off

Risk level: SEVERE

Disables ALL CPU vulnerability protections: Spectre v1/v2, Meltdown, MDS, L1TF, Retbleed, SRBDS, and more. Your CPU becomes vulnerable to side‑channel attacks that can leak data between processes.

Safe if: This is a dedicated gaming rig used by one person, running only trusted software from known sources (Steam, official repos). The attack surface is minimal on a single‑user gaming desktop.

Do NOT use if: The machine handles sensitive data, runs multi‑tenant workloads, executes untrusted binaries, or serves as a server accessible from the network.

Practical take: Benchmark your games with and without mitigations=off. If you see a <5% gain, it's probably not worth the risk. If you see 10‑15%, decide for yourself.

Power/Thermal Warning: intel_idle.max_cstate=1

Risk level: MODERATE

Prevents the CPU from entering deep sleep states (C2–C6). Eliminates wake‑from‑idle latency but increases idle power consumption by roughly 5–15 W depending on silicon. On a desktop with adequate cooling, the thermal impact is negligible. On a laptop, this will noticeably reduce battery life.

Kernel Update Survival Guide

Kernel updates (via pacman -Syu) can reset module parameters or change scheduler availability. After an update:

# 1. Check your NVIDIA module options survived
cat /proc/driver/nvidia/params | grep -E "UsePageAttributeTable|InitializeSystemMemoryAllocations"

# 2. Check if ADIOS is still available (if you switched to it)
cat /sys/block/nvme0n1/queue/scheduler
# If ADIOS is gone, revert to kyber or mq-deadline

# 3. Regenerate initramfs if limine changed
sudo limine-mkinitcpio

Applying

# Add your parameters to /etc/default/limine
sudo nano /etc/default/limine

# Example line:
# KERNEL_CMDLINE[default]+="quiet mitigations=off ... rw root=UUID=..."

# Rebuild boot entries
sudo limine-mkinitcpio

# Reboot to apply
sudo reboot

4. CPU Tuning

Lock the governor to Performance

The performance governor prevents the CPU from spending time in lower P‑states during load, reducing frequency ramp‑up latency. Three mechanisms should be used together, each covers a gap in the others:

Step 1: cpupower service

sudo systemctl enable --now cpupower

Create /etc/cpupower.conf if it doesn't exist:

governor="performance"

Step 2: tmpfiles.d (race‑condition‑proof)

These files apply before any desktop session starts, preventing power manager daemons from overriding the governor:

Create /etc/tmpfiles.d/force-performance.conf:

w /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor - - - - performance

Create /etc/tmpfiles.d/force-epp-performance.conf:

w /sys/devices/system/cpu/cpu*/cpufreq/energy_performance_preference - - - - performance

Why tmpfiles.d? systemd-tmpfiles-setup runs at early boot, before udev fully settles and before any user session. If a power‑profiles‑daemon or similar tries to change the governor later, you've already won the race.

Energy Performance Preference

On Intel CPUs with intel_pstate=active (set in boot parameters), the Energy Performance Preference controls how aggressively the CPU pursues higher P‑states. Locking it to performance means the CPU holds its maximum frequency whenever there's any load.

Verification

cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
# → performance

cat /sys/devices/system/cpu/cpu0/cpufreq/energy_performance_preference
# → performance

5. sched_ext: Swappable Schedulers

#5-sched_ext-swappable-schedulers

linux-cachyos ships EEVDF as its default scheduler, not BORE. See the correction in section 1. EEVDF vs BORE matters less than what actually sets CachyOS apart here: CONFIG_SCHED_CLASS_EXT=y is built into the default kernel, so sched_ext (SCX) is available from install. SCX schedulers are CPU schedulers implemented as loadable BPF programs. You can swap the whole scheduling algorithm at runtime. No reboot, no separate kernel, and you're back on EEVDF the moment you stop one.

Install

#install

sudo pacman -S scx-scheds scx-tools
# scx_loader (background service, D-Bus API) and scxctl (CLI client) come with scx-tools

If you want the newest experimental schedulers before they land in the stable branch, use scx-scheds-git scx-tools-git instead.

Launching one

#launching-one

# One-off test. Ctrl+C stops it and you're back on EEVDF.
sudo scx_bpfland

# Managed via scxctl, which talks to the scx_loader D-Bus service
sudo systemctl enable --now scx_loader
scxctl start --sched bpfland --mode gaming
scxctl get                     # what's running right now
scxctl stop                    # back to EEVDF

CachyOS Kernel Manager and the standalone SCX Manager GUI both wrap the same scx_loader service if you'd rather click through it than type it.

Which scheduler for this hardware

#which-scheduler-for-this-hardware

The i5-9400F is a monolithic 6-core, 6-thread die. No SMT, no heterogeneous P/E cores. That rules out most of the SMT-contention and hybrid-core-aware flags other schedulers expose, so the choice really comes down to raw interactivity under load.

Scheduler	Best for	Notes
`scx_bpfland`	General desktop and gaming	vruntime-based, prioritizes interactive tasks, L2/L3-cache aware placement. Low Latency mode: `-m performance -w`
`scx_cake`	Gaming specifically	Classifies threads into 4 tiers by measured runtime (input/audio, render, physics/AI, background compile) so frame-critical threads never queue behind a shader-compile job. `esports` profile drops the quantum to 1ms for competitive FPS titles
`scx_lavd`	Gaming, latency-sensitive workloads	Built for gaming's latency-critical, communication-heavy pattern. `--performance` flag maximizes responsiveness. `--autopower` adjusts power mode through LAVD's Core Compaction logic, which matters more on a laptop than on this 6C/6T desktop chip
`scx_flash`	Consistency under stress	EDF-based, rewards tasks that yield the CPU early. Gaming mode: `-m all`
`scx_rusty`	Desktop and multimedia	The original CachyOS SCX scheduler pitch. Interactivity close to BORE without a kernel rebuild

What actually won on this rig

#what-actually-won-on-this-rig

Numbers below are from a real CS2 benchmark run on this exact machine (COSMIC/Wayland, i5-9400F, GTX 1650 at the 75W/1860MHz lock from section 9), launch options LOW_LATENCY_LAYER=1 LOW_LATENCY_LAYER_REFLEX=1 SDL_VIDEODRIVER=wayland gamemoderun %command% -sdlaudiodriver pulse -novid, both schedulers on their low-latency/esports profile:

CS2, scx_cake vs scx_lavd (lowlatency, 2 runs)
           Avg FPS   P1 low   P99 (ms)
cake r1      208.9    116.7      8.57
cake r2      208.1    109.1      9.16
cake avg     208.5    112.9      8.87
lavd r1      208.6    108.2      9.24
lavd r2      200.4     91.2     10.96
lavd avg     204.5     99.7     10.10

scx_cake beat scx_lavd on this box, mainly on consistency: cake's 1% low stayed within about 8ms of itself across both runs, lavd's second run dropped a spike that cost it roughly 13ms of 1% low margin against cake's average. That's one bad run out of two though, not a verdict. Worth another 2 to 3 passes before writing lavd off entirely, especially since it's tuned around gaming workloads on paper and might just need a different flag set on a 6-core chip with no SMT than it gets on the hardware it's usually benchmarked on. scx_bpfland and scx_cake were the best performers overall on this setup. Your mileage may genuinely vary since SCX performance is workload- and CPU-topology-dependent. Run your own comparison with cachyos-benchmarker and the games you actually play before locking anything in.

Autostart on boot

#autostart-on-boot

sudo mkdir -p /etc/scx_loader
sudo cp /usr/share/scx_loader/config.toml /etc/scx_loader/config.toml

# /etc/scx_loader/config.toml
default_sched = "scx_cake"
default_mode = "Gaming"

sudo systemctl enable --now scx_loader.service

Swap to bpfland for testing with scxctl switch --sched bpfland --mode gaming, no need to edit the config file for a quick comparison.

power-profiles-daemon integration: scx_loader can auto-switch scheduler profiles when you change the system power profile (Performance to Gaming mode, and so on) through a CachyOS patch to power-profiles-daemon. This guide masks power-profiles-daemon in section 14 so it stops fighting the locked-performance governor. If you've masked it, that auto-switch integration won't fire. Set default_mode = "Gaming" directly instead, or use scx_loader's game-performance hook so it switches automatically when Gamemode launches a game, regardless of what the power profile daemon is doing.

ananicy-cpp compatibility

#ananicy-cpp-compatibility

Section 10 tells you to disable ananicy-cpp because it fights Gamemode's renice values. That's still correct and has nothing to do with sched_ext. Separately, ananicy-cpp used to cause stalls when combined with early SCX schedulers. On current scx-scheds releases the two are generally fine together. If you re-enable ananicy-cpp for other reasons and see stalls, disabling it is still the first thing to try.

Benchmarking before you commit

#benchmarking-before-you-commit

sudo pacman -S schbench cachyos-benchmarker
schbench -m 2 -t 8 -r 60          # wakeup and request latency percentiles
cachyos-benchmarker ~/bench/      # full suite: compiles, encodes, hashes, renders, per-scheduler

Run the same test across EEVDF (default, nothing loaded), BORE if you install that variant, and 2 to 3 SCX schedulers before picking one permanently. A scheduler that wins on paper for "gaming" can still lose on your specific combination of a 6-core no-SMT budget CPU and one GPU doing all the rendering work. That's exactly what the cake vs lavd numbers above show: run it yourself, don't take a table off a forum post as gospel, including this one.

6. Memory & Swap

ZRAM compression: consider lz4 over zstd

CachyOS defaults to zstd compression for ZRAM. On a CPU with limited threads (6 cores, no hyperthreading), zstd compression can consume measurable CPU time during heavy memory pressure.

Recommendation: Switch to lz4 if you have a 6‑core CPU without hyperthreading or if you notice CPU spikes during memory pressure.

Create /etc/systemd/zram-generator.conf.d/override.conf:

[zram0]
compression-algorithm = lz4

lz4: ~3–4× faster compression, ~5× faster decompression, ~15–20% worse compression ratio
zstd (default): Better compression, more CPU usage
Tradeoff: lz4 uses slightly more RAM for compressed pages but saves CPU. With 16 GB RAM, the extra RAM cost is negligible.

Apply:

sudo systemctl restart systemd-zram-setup@zram0

Swappiness

CachyOS automatically sets vm.swappiness = 150 via a udev rule when ZRAM activates. This high value tells the kernel to prefer swapping cold anonymous pages into ZRAM (where they're compressed in RAM) rather than dropping file cache. This is appropriate for ZRAM‑backed swap.

If you want even more aggressive ZRAM usage, set it higher in sysctl:

vm.swappiness = 180

The udev rule overrides to 150 at runtime regardless, the higher value in sysctl is an expression of intent. The effective value will be 150.

Transparent Hugepages

CachyOS already configures this optimally. Verify:

cat /sys/kernel/mm/transparent_hugepage/enabled
# → [always] madvise never

cat /sys/kernel/mm/transparent_hugepage/defrag
# → always defer [defer+madvise] madvise never

Dirty Page Writeback: Fix desktop stuttering during large transfers

The problem: Linux defaults allow up to 20% of RAM to become "dirty" (unwritten data) before throttling writes. On a 16 GB system, that's 3.2 GB. When you download or extract a large file, the kernel buffers everything in page cache, then flushes it all at once. The resulting I/O storm starves the desktop compositor, causing visible stutter and unresponsive UI.

The fix: Force dirty pages to be written out earlier and more gradually.

Add to your sysctl config:

vm.dirty_background_ratio = 3
vm.dirty_ratio = 8

Parameter	Kernel default	Recommended	Effect
`vm.dirty_background_ratio`	10	3	Start writing dirty pages at 3% of RAM dirty (~500 MB on 16 GB)
`vm.dirty_ratio`	20	8	Throttle processes generating writes at 8% dirty (~1.3 GB)

On systems with more RAM, you may want even lower values. The goal is to keep the "dirty burst" small enough that the SSD can flush it without starving other I/O.

7. Storage & Filesystems

XFS mount options

These mount options reduce metadata writes and improve throughput:

For your root filesystem (SSD):

UUID=... / xfs defaults,lazytime,noatime,inode64,logbsize=256k,noquota 0 1

For bulk HDD storage:

UUID=... /mount/point xfs defaults,noatime,nofail,allocsize=64m 0 0

Option	Effect	When to use
`lazytime`	Buffer atime/mtime/ctime in memory, flush opportunistically	Always on SSD root, massive metadata write reduction
`noatime`	Never update access time	Always, there is almost no reason to track atime on a desktop
`inode64`	Allow inodes across full 64‑bit space	Required for volumes >1 TB, harmless otherwise
`logbsize=256k`	Larger XFS journal buffer	SSD‑only, improves metadata throughput. Default is 32k
`noquota`	Disable quota accounting	Only if you don't need filesystem quotas
`allocsize=64m`	Pre‑allocate 64 MB extents	HDD only, reduces fragmentation for large sequential writes (media, backups, Steam downloads)
`nofail`	Boot continues if drive missing	Use for non‑critical external/secondary drives

tmpfs on /tmp

tmpfs /tmp tmpfs defaults,noatime,mode=1777 0 0

Places /tmp in RAM, reduces SSD writes and speeds up applications that use temporary files. On a 16 GB system, the typical /tmp usage is negligible compared to total RAM.

F2FS for secondary/scratch SSDs

If you have a second SSD used for downloads, extraction scratch space, or disposable data, consider formatting it as F2FS instead of ext4/XFS. F2FS is a flash‑optimized filesystem that excels at sequential write workloads. It's ideal for drives where data integrity isn't critical (temp downloads, game library mirroring).

# Format (DESTRUCTIVE: wipes the drive)
sudo mkfs.f2fs -f /dev/sdX

# Mount
UUID=... /mount/point f2fs defaults,noatime,nofail 0 0

IO Scheduler: switch to ADIOS

CachyOS defaults are mq-deadline (SATA SSD), kyber (NVMe), and bfq (HDD). For a single‑user desktop with mixed workloads, ADIOS (Adaptive Disk I/O Scheduler, a CachyOS kernel feature) provides better all‑around performance.

ADIOS uses per‑CPU dispatch queues with adaptive batching. It balances throughput and latency better than mq-deadline (which can hurt latency with aggressive merges) and kyber (which targets fixed latency at the cost of throughput).

Create /etc/udev/rules.d/60-ioschedulers.rules:

# HDD: BFQ remains the right choice
ACTION=="add|change", KERNEL=="sd[a-z]*", ATTR{queue/rotational}=="1", \
    ATTR{queue/scheduler}="bfq"

# SSD: use ADIOS instead of mq-deadline
ACTION=="add|change", KERNEL=="sd[a-z]*|mmcblk[0-9]*", ATTR{queue/rotational}=="0", \
    ATTR{queue/scheduler}="adios"

# NVMe: use ADIOS instead of kyber
ACTION=="add|change", KERNEL=="nvme[0-9]*", ATTR{queue/rotational}=="0", \
    ATTR{queue/scheduler}="adios"

Apply:

sudo udevadm control --reload-rules && sudo udevadm trigger

Verify:

cat /sys/block/sdX/queue/scheduler
# Should show [adios] for SSDs, [bfq] for HDDs

NVMe Power Management (Optional)

Some NVMe drives aggressively enter power‑saving states, causing micro‑stutters when they wake up.

# Disable APST (Autonomous Power State Transition)
echo 0 | sudo tee /sys/class/nvme/nvme0/device/power/control
# To make it permanent, add to a systemd service or rc.local

8. GPU: NVIDIA

Module parameters (CachyOS provides these)

You don't need to change these, they're already set by CachyOS at /usr/lib/modprobe.d/nvidia.conf:

options nvidia NVreg_UsePageAttributeTable=1
    NVreg_InitializeSystemMemoryAllocations=0
    NVreg_DynamicPowerManagement=0x02

Enable persistence daemon

Keeps the NVIDIA kernel driver loaded even with no applications using the GPU. Eliminates the ~1–3 second driver initialization delay when launching the first GPU application after boot.

sudo systemctl enable --now nvidia-persistenced

Force Full RGB over HDMI

NVIDIA defaults to Limited Range (16–235) over HDMI, causing black crush and washed‑out colors on TVs and monitors. Force Full Range (0–255) RGB.

Create /etc/X11/xorg.conf.d/20-nvidia-full-rgb.conf:

Section "Device"
    Identifier "NVIDIA"
    Driver "nvidia"
    Option "ColorSpace" "RGB"
    Option "ColorRange" "Full"
EndSection

Critical for TV users, without this, dark scenes in games and movies will look crushed and gray.

Note: If you have nvidia-settings daemon running, it might override these settings. Kill it if you see issues:

sudo pkill nvidia-settings

Environment variables

Set these system‑wide in /etc/environment:

__GL_THREADED_OPTIMIZATION=1
__GL_SHADER_DISK_CACHE=1
__GL_SHADER_DISK_CACHE_SKIP_CLEANUP=1
__GL_VRR_ALLOWED=0
__GL_SYNC_TO_VBLANK=0
__GL_MaxFramesAllowed=1
VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json
WINEDEBUG=-all

Variable	Effect	Skip if
`__GL_THREADED_OPTIMIZATION=1`	Multi‑threaded OpenGL pipeline	,
`__GL_SHADER_DISK_CACHE=1`	Persistent compiled on disk	If disk space is extremely tight
`__GL_SHADER_DISK_CACHE_SKIP_CLEANUP=1`	Stop driver from purging cache periodically	,
`__GL_VRR_ALLOWED=0`	Disable VRR (GSync/FreeSync)	If you use a VRR display and want tear‑free
`__GL_SYNC_TO_VBLANK=0`	Disable V‑Sync at driver level	If tearing bothers you
`__GL_MaxFramesAllowed=1`	Pre‑rendered frames = 1 (lowest latency)	,
`VK_ICD_FILENAMES`	Force NVIDIA Vulkan ICD	Only if you have multiple GPUs
`WINEDEBUG=-all`	Suppress Wine debug output	If you're debugging Wine issues

Shader Cache Size

__GL_SHADER_DISK_CACHE=1 enables the cache, but without a size limit, the driver uses its default cap of 1 GB (increased from 128 MB in driver 460). Modern titles (especially DXVK/VKD3D) can generate 2–4 GB of shaders across your library. When the cache exceeds the limit, the driver purges older entries, forcing recompilation on next launch, causing stutter and longer load times.

Set a global size limit to prevent eviction:

mkdir -p ~/.config/environment.d
echo "__GL_SHADER_DISK_CACHE_SIZE=12000000000" > ~/.config/environment.d/shader_cache.conf

Log out and back in, or reboot. Verify with:

env | grep __GL_SHADER_DISK_CACHE_SIZE

Note: Setting __GL_SHADER_DISK_CACHE_SIZE globally ensures every application uses the same limit. If you set it per‑game and omit it on some launches, the driver reverts to the 1 GB default and wipes the cache when it exceeds the smaller limit.

Optional, clear legacy caches (old .nv path still uses 128 MB default):

rm -rf ~/.nv/GLCache ~/.cache/nvidia/GLCache

The driver will rebuild under the new XDG path (~/.cache/nvidia/GLCache) with the 4 GB cap applied.

V‑Sync and VRR Warning

Risk level: MINOR (visual)

__GL_SYNC_TO_VBLANK=0 and __GL_VRR_ALLOWED=0 disable all synchronization. You will see screen tearing, especially below your display's refresh rate. This is a latency vs. visual quality tradeoff, uncapped, unsynchronized frames produce the lowest possible input latency.

Set both to 1 if you prefer tear‑free output and don't need the absolute minimum latency.

NVIDIA Wayland Reality Check

Wayland on NVIDIA is finicky. The guide assumes nvidia_drm.modeset=1 is enough. It's necessary, but not sufficient.

Requirements for a smooth Wayland experience:

NVIDIA driver ≥ 550.xx (check with nvidia-smi --query-gpu=driver_version --format=csv,noheader)
Compositor with explicit sync support:
KDE Plasma ≥ 6.1
GNOME ≥ 46
COSMIC (latest)
If you get a black screen or flickering, switch to X11 (nvidia_drm.modeset=0 or just choose X11 session at login).

9. GPU Overclocking via LACT

LACT controls NVIDIA GPUs through the NVML API. It provides a GUI for overclocking, fan curves, and monitoring.

sudo systemctl enable --now lactd

The Clock Lock Strategy (Not "Under‑volting")

Modern NVIDIA GPUs (Turing and later) don't support traditional undervolting in Linux. Instead, use a clock lock: lock the GPU clock below the stock boost target and add a positive offset to the P8 voltage point.

Why this works: The GPU runs at a higher voltage than it naturally would at that clock, eliminating power throttling while staying cooler than stock boost behavior.

For GTX 1650 / Turing‑class GPUs:

Setting	Value	Notes
Power Limit	Max (75W for GTX 1650)	Your GPU's maximum, check with `nvidia-smi -q -d POWER`
GPU Min Clock	300 MHz	Stock minimum
GPU Max Clock	~100 MHz below stock boost	For GTX 1650: 1860 MHz (stock boost ~1905–1935 MHz)
GPU Offset (P8)	+100 MHz	Raises voltage at the locked clock
VRAM Offset (P8)	+500 to +900 MHz	Start at +500, test for stability, increase gradually

Finding your stock boost clock:

nvidia-smi -q -d CLOCK | grep "Max Clocks"

For other GPUs: Lock your max clock ~50–100 MHz below the stock boost target. The offset depends on silicon quality, start at +50 MHz and test.

Fan curve

Aim for quiet at idle‑to‑moderate loads with aggressive ramp past 70°C:

Temp	40°C	45°C	60°C	70°C	80°C
Fan	40%	50%	55%	65%	80%

Adjust based on your card's cooler. The goal is to stay under 75°C during extended loads.

Stability testing

Run a demanding game or GPU benchmark for at least 2–3 hours. If stable, run your longest typical gaming session (8+ hours). If you see artifacts, crashes, or nvidia-smi reporting errors, reduce the VRAM offset by 50 MHz and retest.

10. Gaming Pipeline

Gamemode

Gamemode is Feral Interactive's performance daemon. Launch games with gamemoderun %command% to apply optimizations.

Recommended config at /etc/gamemode.ini:

[general]
renice = 20
ioprio = 0

[gpu]
apply = 1
nv_powermizer_mode = 1

[custom]
start = nvidia-smi -pm 1
end = nvidia-smi -pm 0

Setting	Value	What it does
`renice`	20	Highest scheduling priority (lowest nice value)
`ioprio`	0	Realtime I/O priority
`nv_powermizer_mode`	1	Force GPU to maximum performance clock level
`start`/`end` scripts	Persistence mode toggle	Enables driver persistence on game launch, disables on exit

Gamemode group

Create /etc/security/limits.d/10-gamemode.conf:

@gamemode - nice -10

Add any user who games to the gamemode group:

sudo usermod -aG gamemode $USER

This allows renice to -10 without root. Gamemode itself re‑nices further (to the equivalent of -20) via its daemon privileges.

Conflict: disable ananicy‑cpp

CachyOS ships with ananicy-cpp, an automatic process priority daemon. Disable it if you use Gamemode. The two fight over nice values, ananicy‑cpp periodically rescans and resets priorities, causing processes to bounce between levels. This manifests as microstuttering in games.

sudo systemctl disable --now ananicy-cpp

Steam launch options

Native Linux games:

SDL_VIDEODRIVER=wayland gamemoderun %command%

Windows games via Proton:

PROTON_ENABLE_WAYLAND=1 gamemoderun %command%

Forcing fullscreen to a specific display (e.g., TV on HDMI):

PROTON_ENABLE_WAYLAND=1 SDL_VIDEO_FULLSCREEN_DISPLAY=HDMI-A-1 gamemoderun %command%

Find your display name with:

echo $DISPLAY
# Or for Wayland: check your compositor's display settings

Texture quality overrides

These replicate NVIDIA Control Panel "High Quality" texture filtering at the application level, useful since NVIDIA Settings on Linux doesn't expose these controls per‑game.

DXVK (DirectX 9/10/11 games via Proton):

DXVK_ANISO=16 DXVK_LODBIAS=-0.5 PROTON_ENABLE_WAYLAND=1 gamemoderun %command%

Native OpenGL:

__GL_LOG_ANISO=16 __GL_TEXTURE_LOD_BIAS=-0.5 SDL_VIDEODRIVER=wayland gamemoderun %command%

Variable	Effect
`DXVK_ANISO=16` / `__GL_LOG_ANISO=16`	16× anisotropic filtering, sharpens textures at oblique angles
`DXVK_LODBIAS=-0.5` / `__GL_TEXTURE_LOD_BIAS=-0.5`	Negative LOD bias, forces higher‑resolution mipmaps

Gamescope: probably skip it

Gamescope is Valve's microcompositor. It adds a compositing layer between the game and your display. On a native Wayland compositor that already supports fullscreen bypass (COSMIC, KWin, Mutter), Gamescope adds input latency without benefit. Only use it if you need its specific features (FSR upscaling, HDR tonemapping, nested sessions).

Proton

Install the CachyOS Proton builds for additional performance patches:

sudo pacman -S proton-cachyos-native proton-cachyos-slr

These include scheduler optimizations and Wine patches beyond Valve's upstream Proton.

MangoHud

Minimal, data‑dense overlay config at ~/.config/MangoHud/MangoHud.conf:

hud_compact
hud_no_margin
text_outline
background_alpha=0.3
font_size=14
table_columns=2

fps
frametime

gpu_stats
gpu_temp
gpu_core_clock
gpu_mem_clock
vram

cpu_stats
cpu_temp
cpu_mhz

wine

Key design choices:

No frametime graph, it consumes massive vertical space for information you can get from the frametime number
Two‑column layout (table_columns=2), keeps the overlay compact
Compact mode, eliminates padding between rows
Wine metrics, shows DXVK/VKD3D version and other Proton‑relevant info

OptiScaler

For DLSS‑style upscaling on non‑RTX hardware (GTX 1650, 1660, etc.), OptiScaler injects FSR/XeSS into games that only support DLSS. Install as a Wine DLL override and configure per‑game.

OnlineFix / cracked game support

For games using OnlineFix multiplayer patches, add these Wine DLL overrides:

onlinefix64=n;winmm=n,b;steam_api64=n;version=n,b

Reference: onlinefix-linux

11. Audio: PipeWire Low Latency

Quantum tuning

Lower the PipeWire quantum (processing block size) for reduced audio latency. Create /etc/pipewire/pipewire.conf.d/10-low-latency.conf:

context.properties = {
    default.clock.rate            = 48000
    default.clock.quantum         = 128
    default.clock.min-quantum     = 128
    default.clock.max-quantum     = 512
}

Setting	Value	Effect
`clock.rate`	48000 Hz	Standard sample rate
`quantum`	128 samples	~2.67 ms processing cycle
`min-quantum`	128	Minimum allowed block size
`max-quantum`	512	Maximum allowed block size (~10.67 ms)

Latency: 128 samples / 48000 Hz ≈ 2.67 ms per cycle. Round‑trip ≈ 2× this (~5.3 ms) plus hardware conversion.

Tradeoff: Lower quantum = lower latency but higher CPU usage for audio processing. 128 samples is appropriate for competitive gaming and rhythm games. For general desktop use (media playback, voice chat, casual gaming), 256 samples (~5.3 ms) is sufficient and uses less CPU.

Apply:

systemctl --user restart pipewire

Verify:

pw-top
# Look at the QUANT column during playback

CachyOS audio stack

CachyOS provides these out of the box (you don't need to configure them):

rtkit, grants realtime scheduling to PipeWire
Audio group realtime priority, @audio - rtprio 99
CPU DMA latency device, audio group can suppress CPU C‑states
snd_hda_intel power saving disabled on AC, prevents audio crackling
HPET/RTC permissions, audio group access to high‑precision timers

12. Network

Sysctl additions

Add these to your custom sysctl config (e.g., /etc/sysctl.d/99-latency.conf):

net.core.netdev_max_backlog = 16384
net.ipv4.tcp_fastopen = 3
net.core.somaxconn = 1024

Setting	CachyOS default	Recommended	Why
`netdev_max_backlog`	4096	16384	Larger per‑CPU packet buffer, prevents drops under burst traffic (game downloads, streaming)
`tcp_fastopen`	0 (off)	3	TCP Fast Open (client + server), saves 1 RTT on repeat connections
`somaxconn`	4096 (kernel default)	1024	Tuned for desktop workloads, not servers. Prevents excessive queuing

CachyOS defaults you shouldn't change

Queue discipline: fq_codel, actively fights bufferbloat. Don't change this.
TCP congestion: cubic, standard and well‑tested for Internet use.
TCP keepalive: 120 seconds, detects dead connections faster than kernel default.
DNS: systemd-resolved with DNSSEC validation.
NTP: Cloudflare primary, Google fallback.

13. System Limits & Sysctl

File descriptor limits

Wine and Proton open many file descriptors simultaneously. The CachyOS default (2,097,152 for system services, 1,048,576 for user sessions) is generous, but some games benefit from higher limits.

In /etc/security/limits.conf:

* hard core 0
* soft core 0
* hard nofile 524288
* soft nofile 524288
root hard nofile 524288
root soft nofile 524288

Setting	Value	Why
Core dumps	0 (disabled)	Prevents crash dumps from consuming disk space. Re‑enable if you're debugging crashes.
nofile	524288	Higher than CachyOS user default (1M), appropriate for Wine/Proton

Full recommended sysctl

Create /etc/sysctl.d/99-latency.conf:

vm.swappiness = 180
vm.dirty_background_ratio = 3
vm.dirty_ratio = 8
net.core.somaxconn = 1024
net.core.netdev_max_backlog = 16384
net.ipv4.tcp_fastopen = 3
fs.file-max = 2147483647
fs.inotify.max_user_watches = 524288
fs.inotify.max_user_instances = 1024

Note on swappiness: You set it to 180, but CachyOS ZRAM udev rule overrides to 150 at runtime. The high value signals intent, aggressively push cold anonymous pages to ZRAM. The effective runtime value (150) is already 2.5× the kernel default of 60.

Apply:

sudo sysctl --system

14. Services to Disable

These services are enabled by default on CachyOS but are unnecessary for a dedicated gaming desktop. Disabling them reduces background CPU usage, memory consumption, and boot time.

Service	How	Why
`ananicy-cpp`	`sudo systemctl disable --now ananicy-cpp`	Conflicts with Gamemode, causes microstuttering
`power-profiles-daemon`	`sudo systemctl mask power-profiles-daemon`	Overrides manual CPU governor settings
`NetworkManager-wait-online`	`sudo systemctl disable NetworkManager-wait-online`	Unnecessary boot delay on desktop
`lvm2-monitor`	`sudo systemctl mask lvm2-monitor`	Only needed if using LVM volumes
`bluetooth`	`sudo systemctl disable --now bluetooth`	Disable if you don't use Bluetooth peripherals
`geoclue`	`sudo systemctl disable --now geoclue`	Location services, not needed on desktop
`ModemManager`	`sudo systemctl disable --now ModemManager`	Cellular modem management, irrelevant without a modem
`upower`	`sudo systemctl disable --now upower`	Battery monitoring, irrelevant on desktop without UPS

Disable vs mask:

disable, service won't start automatically but can be started as a dependency
mask, service cannot be started at all, even as a dependency of another service

Mask only when you're certain the service will never be needed.

15. Desktop & Wayland

Choosing a compositor

A Wayland‑native compositor that supports fullscreen bypass (direct scanout) is critical for gaming performance. When a game goes fullscreen, the compositor hands the display buffer directly to the GPU, eliminating the compositing step and its associated latency.

Good options:

COSMIC (System76, Rust‑based), actively developed, fullscreen bypass supported
KWin (KDE Plasma), mature, fullscreen bypass, extensive VRR support
Mutter (GNOME), fullscreen bypass, VRR support improving

NVIDIA + Wayland

NVIDIA GPUs require DRM kernel modesetting for Wayland. Ensure your boot parameters include:

nvidia_drm.modeset=1

Without this, Wayland sessions will either fall back to software rendering or fail to start.

Reality check: If you have driver < 550.xx or a compositor without explicit sync, Wayland will be a flickery mess. Don't fight it, use X11.

Gamescope

If your compositor already supports fullscreen bypass, you don't need Gamescope. It adds an extra compositing layer and measurable input latency. Only use it for features your compositor lacks (FSR upscaling on older GPUs, HDR tonemapping, nested Steam Deck sessions).

16. Minimum Viable Tuning (The 80/20)

Don't want to apply the entire guide? These 4 things give 90% of the performance gain with 10% of the effort.

#	Change	Command / Config
1	CPU Governor	`sudo systemctl enable --now cpupower` + `/etc/cpupower.conf` with `governor="performance"`
2	ZRAM → lz4 (if 6‑core)	`/etc/systemd/zram-generator.conf.d/override.conf` → `compression-algorithm = lz4`
3	Dirty Ratio	`vm.dirty_background_ratio = 3`, `vm.dirty_ratio = 8` in sysctl
4	Disable ananicy‑cpp	`sudo systemctl disable --now ananicy-cpp`

If you do only these, you'll solve 80% of the stutter and latency issues. The rest is fine‑tuning.

17. Verification Checklist

After applying each optimization, verify it took effect:

# ── Boot parameters ──
cat /proc/cmdline

# ── CPU ──
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
# → performance

cat /sys/devices/system/cpu/cpu0/cpufreq/energy_performance_preference
# → performance

zcat /proc/config.gz | grep CONFIG_PREEMPT
# → CONFIG_PREEMPT=y

zcat /proc/config.gz | grep CONFIG_HZ=
# → CONFIG_HZ=1000

zcat /proc/config.gz | grep CONFIG_SCHED_CLASS_EXT
# → =y

uname -r
# "-bore" in the string = BORE kernel, otherwise EEVDF

scxctl get
# → currently loaded SCX scheduler, or "none" for EEVDF/BORE

# ── Memory ──
zramctl
# Check compression algorithm and size

sysctl vm.swappiness
# → 150 (CachyOS ZRAM default)

cat /sys/kernel/mm/transparent_hugepage/enabled
# → [always] madvise never

sysctl vm.dirty_background_ratio vm.dirty_ratio
# → 3 and 8

# ── Storage ──
cat /sys/block/sdX/queue/scheduler
# → [adios] for SSDs, [bfq] for HDDs

# ── GPU ──
nvidia-smi -q | grep Persistence
# → Persistence Mode: Enabled

nvidia-smi -q -d CLOCK | grep "Max Clocks"
# → Check your overclock values

# ── Audio ──
pw-top
# Check QUANT column during playback (should show 128)

# ── Network ──
sysctl net.ipv4.tcp_fastopen
# → 3

sysctl net.core.netdev_max_backlog
# → 16384

# ── File limits ──
ulimit -n
# → 524288

# ── Gamemode (run while a game is active) ──
gamemoded -s
# → "gamemode is active"

# ── Disabled services ──
systemctl is-active ananicy-cpp power-profiles-daemon
# → inactive

18. System Recovery (If You Break It)

If your system doesn't boot after changing kernel parameters:

Boot from USB (CachyOS installer ISO).
Mount your root partition:

mount /dev/sdXY /mnt
mount /dev/sdXZ /mnt/boot  # if separate

Chroot in:

arch-chroot /mnt

Revert your changes:

# Restore limine backup if you made one
cp /etc/default/limine.bak /etc/default/limine
# Or simply remove the offending parameter
sudo limine-mkinitcpio

Reboot (exit chroot, umount, reboot).

Alternatively, if you use BTRFS and made a snapshot (@pre-optimization), you can roll back from the bootloader menu (if CachyOS configured it) or rename the subvolumes from the chroot.

19. Configuration File Reference

Complete copy‑paste‑ready files. All paths are absolute from root unless noted.

/etc/default/limine (kernel command line)

KERNEL_CMDLINE[default]+="quiet mitigations=off nowatchdog nmi_watchdog=0 nvidia_drm.modeset=1 tsc=reliable clocksource=tsc intel_pstate=active preempt=full split_lock_detect=off pcie_aspm=performance intel_idle.max_cstate=1 transparent_hugepage=madvise splash rw root=UUID=..."

Apply: sudo limine-mkinitcpio

/etc/scx_loader/config.toml

default_sched = "scx_cake"
default_mode = "Gaming"

Apply: sudo systemctl enable --now scx_loader.service

/etc/tmpfiles.d/force-performance.conf

w /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor - - - - performance

/etc/tmpfiles.d/force-epp-performance.conf

w /sys/devices/system/cpu/cpu*/cpufreq/energy_performance_preference - - - - performance

/etc/sysctl.d/99-latency.conf

vm.swappiness = 180
vm.dirty_background_ratio = 3
vm.dirty_ratio = 8
net.core.somaxconn = 1024
net.core.netdev_max_backlog = 16384
net.ipv4.tcp_fastopen = 3
fs.file-max = 2147483647
fs.inotify.max_user_watches = 524288
fs.inotify.max_user_instances = 1024

Apply: sudo sysctl --system

/etc/security/limits.conf (append)

* hard core 0
* soft core 0
* hard nofile 524288
* soft nofile 524288
root hard nofile 524288
root soft nofile 524288

/etc/security/limits.d/10-gamemode.conf

@gamemode - nice -10

/etc/environment (append)

__GL_THREADED_OPTIMIZATION=1
__GL_SHADER_DISK_CACHE=1
__GL_SHADER_DISK_CACHE_SKIP_CLEANUP=1
__GL_VRR_ALLOWED=0
__GL_SYNC_TO_VBLANK=0
__GL_MaxFramesAllowed=1
VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json
WINEDEBUG=-all

/etc/systemd/zram-generator.conf.d/override.conf

[zram0]
compression-algorithm = lz4

Apply: sudo systemctl restart systemd-zram-setup@zram0

/etc/udev/rules.d/60-ioschedulers.rules

# HDD: use BFQ
ACTION=="add|change", KERNEL=="sd[a-z]*", ATTR{queue/rotational}=="1", \
    ATTR{queue/scheduler}="bfq"

# SSD: use ADIOS
ACTION=="add|change", KERNEL=="sd[a-z]*|mmcblk[0-9]*", ATTR{queue/rotational}=="0", \
    ATTR{queue/scheduler}="adios"

# NVMe: use ADIOS
ACTION=="add|change", KERNEL=="nvme[0-9]*", ATTR{queue/rotational}=="0", \
    ATTR{queue/scheduler}="adios"

Apply: sudo udevadm control --reload-rules && sudo udevadm trigger

/etc/pipewire/pipewire.conf.d/10-low-latency.conf

context.properties = {
    default.clock.rate            = 48000
    default.clock.quantum         = 128
    default.clock.min-quantum     = 128
    default.clock.max-quantum     = 512
}

Apply: systemctl --user restart pipewire

/etc/gamemode.ini

[general]
renice = 20
ioprio = 0

[gpu]
apply = 1
nv_powermizer_mode = 1

[custom]
start = nvidia-smi -pm 1
end = nvidia-smi -pm 0

/etc/X11/xorg.conf.d/20-nvidia-full-rgb.conf

Section "Device"
    Identifier "NVIDIA"
    Driver "nvidia"
    Option "ColorSpace" "RGB"
    Option "ColorRange" "Full"
EndSection

/etc/fstab (example: replace UUIDs)

UUID=<boot-uuid>  /boot  vfat  defaults,umask=0077  0 2
UUID=<root-uuid>  /      xfs   defaults,lazytime,noatime,inode64,logbsize=256k,noquota  0 1
tmpfs             /tmp   tmpfs defaults,noatime,mode=1777  0 0
UUID=<hdd-uuid>   /mount/point  xfs  defaults,noatime,nofail,allocsize=64m  0 0
UUID=<ssd-uuid>   /mount/point  f2fs  defaults,noatime,nofail  0 0

~/.config/MangoHud/MangoHud.conf

hud_compact
hud_no_margin
text_outline
background_alpha=0.3
font_size=14
table_columns=2

fps
frametime

gpu_stats
gpu_temp
gpu_core_clock
gpu_mem_clock
vram

cpu_stats
cpu_temp
cpu_mhz

wine

Name		Name	Last commit message	Last commit date
Latest commit History 16 Commits
README.md		README.md
index.html		index.html

Folders and files

Latest commit

History

Repository files navigation

CachyOS Performance Optimization Guide

Table of Contents

1. What CachyOS Already Does

Kernel

Sysctl defaults

ZRAM

Transparent Hugepages

IO Schedulers

NVIDIA Module Options

Systemd Tuning

Other Built‑in Optimizations

2. Before You Begin

3. Boot Parameters

Recommended command line

Parameter reference

Security Warning: mitigations=off

Power/Thermal Warning: intel_idle.max_cstate=1

Kernel Update Survival Guide

Applying

4. CPU Tuning

Lock the governor to Performance

Energy Performance Preference

Verification

5. sched_ext: Swappable Schedulers

Install

Launching one

Which scheduler for this hardware

What actually won on this rig

Autostart on boot

ananicy-cpp compatibility

Benchmarking before you commit

6. Memory & Swap

ZRAM compression: consider lz4 over zstd

Swappiness

Transparent Hugepages

Dirty Page Writeback: Fix desktop stuttering during large transfers

7. Storage & Filesystems

XFS mount options

tmpfs on /tmp

F2FS for secondary/scratch SSDs

IO Scheduler: switch to ADIOS

NVMe Power Management (Optional)

8. GPU: NVIDIA

Module parameters (CachyOS provides these)

Enable persistence daemon

Force Full RGB over HDMI

Environment variables

Shader Cache Size

V‑Sync and VRR Warning

NVIDIA Wayland Reality Check

9. GPU Overclocking via LACT

The Clock Lock Strategy (Not "Under‑volting")

Fan curve

Stability testing

10. Gaming Pipeline

Gamemode

Gamemode group

Conflict: disable ananicy‑cpp

Steam launch options

Texture quality overrides

Gamescope: probably skip it

Proton

MangoHud

OptiScaler

OnlineFix / cracked game support

11. Audio: PipeWire Low Latency

Quantum tuning

CachyOS audio stack

12. Network

Sysctl additions

CachyOS defaults you shouldn't change

13. System Limits & Sysctl

File descriptor limits

Full recommended sysctl

14. Services to Disable

15. Desktop & Wayland