Input Latency Insights

[soul4kills]

What makes solving input latency so hard is that 1 change/tweak alone will not produce a positive result. You have to solve a combination of these things to be able to see a positive result. So you might test 1 tweak out, and come to the conclusion that it didn't work, but in reality that's not the case, because you didn't solve all the other problems that is preventing you from seeing an improvement.

If you are CPU bound, good luck with your latency, because you won't fix it in this situation. I went from a 3600x to 5900x.

Lets understand ALL the barriers that lead to input latency.

Kernal Timer - It's the last step in the chain of requests to CPU Execution. Pretty much dictates over all system latency because it is the last part in the chain. (ISLC tool helps greatly with this)

Timer Resolution - Is the system timer which directly affects Kernal Timer (Windows 11 default is 15+ms). It indirectly affects ISR & DPC latency. Dictates the minimum time it takes for drivers/apps/devices requests to get processed, regardless of your ISR & DPC latency.

ISR & DPC Latency (Accumlative to Kernel Latency) - High latency is usually caused by an accumulation of requests from bad drivers/devices. Timer resolution does not directly affect it, drivers do. Latency is dictated by the operation requested. If it's high, it gets redirected to DPC.
- Example of the problem. Nvidia drivers are not all equal. I'm currently stuck on 576.52, as it's the driver that has the lowest latency. Every new driver that comes out, I test it out and use tools like DPC Latency Checker & LatencyMon. Driver 580.88 looked promising but It still didn't feel as good as 576.52.

Understanding Chain of events from User Input to CPU execution
User Input -> ISR(Higher Priority than DPC, redirects longer requests to DPC)-> Kernel -> CPU Executes < - Kernel <- DPC (Requests gets processed after all ISR requests get processed)

Multiple Layers between communication of Input Device to CPU exectution. These layers are accumulative effectively increasing latency with each layer.
- IO that routes through to the chipset is slower than IO that routes directly to the CPU. So make sure mouse and keyboard are connected to ports directly routed to CPU for lowest latency. GPU should be connected to slot closest to the CPU which is the top most slot.
- SVM & IOMMU in bios effectively adds another layer between user input -> cpu. Core isolation in Windows Defender is an example of this which protects the kernel from malicious drivers. Drivers communication goes through core isolation before being processed by the CPU.
- Hardware can introduce another layer and increase latency for example - All-in-one device hubs are notorious at introducing input latency. As it's multiple devices within a device. Communication goes from CPU -> Driver -> Devices Native Controller -> Singular Device -> Another Device if it's a USB Hub

Number of CPU requests - Example would be number of Hardware Devices/Virtual Devices/Drivers/Applications. (Imagine CPU as a teacher in a classroom doing a continuous roll call. The more students there are, the longer it takes to get back to the top of the list) Goal is to reduce this.
- USB ports are a device, or even a Device Hub. (Even if a device isn't connected to a port, your system attempts to communicate with it to check it's status.)
- A PCIE slot is a device itself.
- Reduce background apps as some particular ones will introduce a lot of latency, not all do though
- Virtual Devices Examples. System Speaker, Remote Desktop Device Redirector Bus, Microsoft Kernal Debug Network Adapter.
- I've disabled all unnecessary devices in device manager.
- I've uninstalled all non-system critical & unecessary Windows Store Apps including Widgets/Cortana/Co-pilot
- I've disabled my LAN device, VGA port, and Audio Device in BIOS
- When disabling devices in Device Manager, sorting by connection lets you see what is connected to what and allows you to disable a full tree of devices or smaller branches of the tree. Be cautious and do you research of what you're disabling.

Hardware-Accelration. I disable hardware-accelration in all my apps to reduce impact on my GPU while gaming, my 5900x is capable of picking up that load. It's so I can also browse the internet or watch videos/streams while gaming with no impact to my GPU resources.

System Stability - For example bad overclocks/underclocks will impact your latency as they introduce errors causing a delay in processing requests. Imagine your computer having hiccups.
- I have a 5900x Underclocked with PBO2, -25 Curve, PPT of 95 (For lower temps), Boost of 150hz (Pretty stable)
- I tried -30 Curve, PPT 95, Boost of 200hz, It appeared stable, but my latency monitors showed me otherwise with higher latency

Efficiency & Performance modes (Efficiency Modes impact latency negatively as they request a high timer resolution to be more eco friendly)
- Enabling High Performance mode in NVIDIA helps a lot with latency, as it's requesting a lower timer resolution. (I personally enable it per app, and globally set it to optimal performance)
- Enabling Ultra/High performance mode in windows has the same effect (Dual CCD AMD X3d chips require balanced mode in addition to game mode on so windows knows which CCD to use when gaming)
- I switch to Ultra Performance mode when gaming and switch to power saver when I'm not

Polling Rates
- 1000hz(1ms) is enough, 2000hz(0.5ms) at most. Anything above that, you're just spamming your CPU with requests introducing latency. If your kernel timer is only capable of 0.5ms, 2000hz, there is no improvement going to 8000hz. So understand the tolerances of your setup, and know when going above a certain metric has nothing to offer.

Bad tweaks or settings suggested by tweak guides - Some examples
- disabledynamictick has a negative impact on my latency.
- XMP 1 profile also has a negative effect. Not as stable as XMP 2.
- Disabling High Precision Event Timer in Device Manager has a negative impact. Not only that, it gets worst over time because your system needs another timer to sync with so it doesn't start to desync, leading to inaccurate metrics in performance monitoring apps. HPET is what your system timer periodically syncs with.
- Disabling Windows Hardware-accelerated GPU scheduling & Optimizations for windowed games. Has no effect on latency, if it does for you, your problem may be something else, and turning these off are just a work around to your problem. Because they actually improve gaming in how they were designed. Lots of improvements and no drawbacks with these settings.

Very effective tweaks & tools
- GPU MSI Enabled + High Priority Interrupt (Allows your GPU to communicate more directly with the CPU at a higher priority) only effective if you've solved all your other latency issues) disclamer- has a negative effect on FPS if you're CPU bound.
- ISLC Tool - USE WHILE GAMING (Adjusts Timer Resolution as low as possible (0.50 is recommended)
- DPC Latency Checker.exe - USE WHILE GAMING (Allows monitoring of your Kernal Latency) It has an interesting built in side effect, it effectively requests lowest timer possible from the system at a global level allowing ISLC tool to have an effect. Without this running, whether or not you set your Timer Resolution to 0.5ms, Kernel Timer will always operate at the system default of 15ms, until a correctly configured app requests for the lowest timer possible. That's what this app does.
- Disable overlays, has little impact but if you want to squeeze everything out of latency, this can help

Misconceptions on interpreting latency monitors
- Being idle will result in high latency, as nothing is requesting a lower Timer Resolution, so system timer will default to 15ms
- Actively playing a game will request a lower Timer Resolution, and will result in lower latency indications in the latency monitors
- Just moving a high polling rate mouse quickly can request a lower Timer Resolution

Windows Version
- Windows 11 changed it's Timer Resolution behavior. In windows 10, it used to be a global thing, where everything will operate at the lowest requesting app or device. But in Windows 11, it doesn't do that anymore. It only adjusts at each request and goes back to default of 15ms, making it dynamic. ISLC can change it to windows 10 behavior with "GlobalTimerResolutionRequest".

I'm on windows 11, AMD 5900x underclocked with a RTX 2060 stock. Wireless keyboard and G Pro Wireless mouse. Playing with a intel wifi connection. I prefer Windowed Fullscreen Modes in my games for the convenience of task switching without delay. I also browse the internet and watch videos with very minimal impact to my input latency while gaming. I also have a lot of background apps running. And my input latency is still great. It's between 0.5ms to 1ms. Smooth and snappy. While also having an IPAD as a second monitor with the app "SpaceDesk".

Trust me, my latency used to be bad, took me years to figure it out. And I finally did with the understanding I laid out above.

I decided to write this after reading a few of these forum posts, and felt like I needed to help and give people some needed insight. I'll see a lot of posts talking about a solution by doing this or that, but really it didn't do anything. So much placebo going on.

So that's it, hope this helps. Sorry for the length and the fact that it has no order to it, didn't really plan on writing this. I don't have deep technical knowledge, but I did my best to explain it as best I could, so don't crucify me for getting some things slightly wrong, and offer up your better explanation if possible.