By Walter Jobs, Technology writer and editor · Published 10 October 2026
Most advice on this is a list of twenty things to try, which is how you spend an evening changing settings at random. A frame rate drop has a shape, and the shape tells you which of six causes you are looking at. Work from the shape and you will usually be down to one or two candidates before you change anything.
So this article is organised by symptom rather than by fix. Find the row that matches what you are seeing, then read that section.
| What you are seeing | Most likely cause | How to confirm it |
|---|---|---|
| Fine for the first few minutes, then settles lower and stays there | Thermal or power limit | Watch the temperature and clock speed, not the frame rate |
| A hitch in the same places, the first time you go there, never again | Shader or pipeline compilation | Revisit the area; if it is smooth, that was it |
| Fine alone, bad after alt-tabbing or with a browser open | Video memory budget | Close the other application and see if it recovers |
| Bad for a few minutes, cured itself, you never found out why | A background scan or indexing pass | Check what ran at that time |
| Everything was fine until a Windows change | A security or power setting that moved | Compare against what changed, not against a guide |
| Only on battery, or only after unplugging | Processor power management | Compare clock speeds on and off power |
Shape one: good, then worse, then stable
This is the most common and the easiest to confirm. If a game runs well for two to ten minutes and then settles at a lower frame rate that stays put, you are watching a chip reach a limit and back off. Nothing is broken.
The processor side is documented plainly. Intel’s own description: “When a core exceeds the set throttle temperature, it will reduce power to maintain a safe temperature level.” Intel does not publish a single number for that, and nor should anyone else: “The maximum junction temperature limit varies per product and usually is between 100°C-110°C.” So a universal “Intel CPUs throttle at 100” is wrong in both directions.
AMD states the relationship directly, and this is the sentence to keep: “there is a direct relationship between temperature, power, and performance and when the processor reaches its specified maximum operating temperature (Tjmax), its power and performance would also be at their limit.”
Which brings up the figure that panics people. A current Ryzen part carries a specified maximum operating temperature of 95°C on AMD’s own product page. That is a design ceiling, reached intentionally under sustained load, not a fault. The chip is doing what the specification says it does. What you lose at that ceiling is headroom, not safety, and the mechanism is AMD’s boost behaviour, which it documents as depending on temperature, workload, core count, socket power, current draw and firmware configuration, adjusting “frequencies up to 1000 times per second”. AMD also states the pattern: “light workloads experience the highest boost frequencies, while heavier multi-core and/or sustained workloads are more likely to encounter a limit and receive less boost.”
Note that Tjmax is per part. Look it up on the specific chip’s own manufacturer page rather than applying 95 to everything.
How to confirm: put temperature and clock speed on screen next to the frame rate, with an overlay that reports both, and watch for the clock dropping as the temperature rises. Our guide to on-screen counters covers which overlays carry those metrics. If the clock falls as the temperature climbs, you have your answer and the fix is airflow, paste, fan curves or a cooler, not settings.
One cause AMD names that catches people out: “If the CPU temperature is higher than expected during what appears to be system idle, a background application (such as RGB and other tools) could be the issue since it is constantly polling the CPU.”
Shape two: a hitch in the same place, once
If the stutter is a hard hitch rather than a sag, happens at specific moments like a new weapon or a new area, and does not happen again when you return to that spot, you are watching shader compilation. This is the best-documented cause on the list and the one least often identified correctly.
The Khronos Group, which maintains Vulkan, describes it without hedging. Creating a graphics pipeline “requires compiling VkShaderModule internally. This will have a significant increase in frame time if performed at runtime.” And the consequence, in their words: “Building pipelines dynamically without a pipeline cache can result in a sudden framerate drop.” Their best practices list it under “Don’t”: “Create pipelines at draw time without a pipeline cache (introduces performance stutters).”
Epic’s engine documentation gives the magnitudes, and they are illuminating. Its tooling flags a compile as a hitch past a threshold, and the default is set high on purpose: “A PSO compilation is marked as a hitch if the compilation took longer than a certain amount of milliseconds”, with “the default value of 20 milliseconds is high because the first hits on the driver cache can take a long time”. Its own worked profiling example lists “a few 5 to 10 msec hitches from PSO precaching misses” and “a big one of 117 msec which will be noticeable by the player”. A separate Epic page notes that “generating a new PSO on-demand can take 100 or more milliseconds”.
A 117 millisecond hitch at 60 fps is seven frames missed in a row. That is a visible lurch, not a dip in an average, which is why it never shows up in a frame rate counter and always shows up in your hands.
Two things follow that are worth knowing.
The first is that this mostly fixes itself. Compiled pipelines get cached, by the game and by the graphics driver, so the second visit is smooth. If a game offers a shader pre-compilation step at startup, taking it trades a one-off wait for the hitches.
The second is a warning about testing it, straight from Epic: “Without it, hitches may be masked by the PSO cache built by the graphics driver and left over from the previous runs.” Any test of shader stutter that does not clear the driver cache first is measuring the cache rather than the game. This is why “I tested it and the stutter is gone” is unreliable evidence.
And one caution against over-diagnosing. Epic’s own example, having listed the compilation hitches, adds: “The other big hitches are not coming from PSO compilation.” Not every spike is a shader.
Shape three: fine alone, bad alongside something else
If the drop correlates with another application rather than with time or place, look at video memory. And the thing to understand here is that running out of it is not a gradual slowdown, whatever the mental model says.
Microsoft’s documentation for the current graphics API explains why, starting with the hardware fact: “GPUs do not yet support page-faulting, so applications must commit data into physical memory while the GPU could access it.” There is no graceful paging mechanism to fall back on. What happens instead: “If an application doesn’t stay within its budget, the process will be intermittently frozen to allow other applications to run and/or the creation APIs will return failure.”
Intermittently frozen. Not slowed. And spilling into system memory, which people imagine as the soft landing, is described as a last resort with a sharp edge: “while it is true that kernel can shift some heaps on discrete adapters from video memory to system memory, it does so only as an extreme last resort”, and “overflowing textures into system memory adds more complexity, as the wrong resource in system-memory can severely impact frame rate”.
Now the part that explains why this appears out of nowhere in a game you have played for weeks. The budget is not a fixed number: “The budget can fluctuate noticeably as background processes wake-up and sleep; and fluctuate dramatically when the user switches away to another application.” Also: “Available physical memory can vary considerably depending on what the user is doing in the background (such as running a browser or watching a video).”
So a browser with thirty tabs, a second monitor playing video, or an alt-tab to a hardware-accelerated application can move the ceiling underneath a running game. Nothing about your settings changed. The space available for them did.
Microsoft also notes why this feels like a new problem: the older API would do the paging itself, and “in some cases, Direct3D 11 applications may have been paging resource contents in and out every frame; and it resulted in acceptable frame rates for the user”. The current API moved that responsibility to the game, which is why modern titles can fall off a cliff where old ones sagged.
How to confirm: an overlay that reports video memory use, then close the other application and see whether it recovers. If it does, you have found it, and the options are lowering texture settings, closing the other thing, or more video memory.
Shape four: bad for a few minutes, then gone
A self-resolving episode usually means something else wanted the machine. Microsoft documents the resource cost of its own background work without connecting it to games, and that is as far as the documentation goes, so we will not overstate it.
On antivirus scanning, Microsoft’s own position is that it is worth optimising around: “The scheduled quick-scan optimization reduces performance degradation by avoiding a scheduled quick scan when that scan isn’t necessary”, and a full scan “can take a few hours or days to complete, depending on the amount and type of data that needs to be scanned”, which is why Microsoft says “we generally don’t recommend scheduling full scans”. Microsoft ships a dedicated tool for the problem, which tells you it is real: “If devices running Microsoft Defender Antivirus are experiencing performance issues, you can use the performance analyzer to improve the performance of Microsoft Defender Antivirus.”
On search indexing, Microsoft states that its thorough mode “may use more system resources” and that first-time indexing “can take up to a couple hours to complete”.
What Microsoft does not document anywhere we could find is an interaction between Windows Update, antivirus scanning or indexing and game performance specifically. Guides assert it constantly. The documentation supports “these things use resources” and stops there, so that is what we will claim.
How to confirm: next time it happens, open Task Manager sorted by CPU and disk, and look at what else is working. If you missed it, check your antivirus scan history for a scan that overlaps the time.
Shape five: everything was fine until a Windows change
Two candidates here, and one of them is the subject of a great deal of confident misinformation.
Memory integrity and virtualisation-based security. This has been widely reported as a frame rate cost. Here is what Microsoft actually publishes, which is narrower: “Memory integrity works better with Intel Kabylake and higher processors with Mode-Based Execution Control, and AMD Zen 2 and higher processors with Guest Mode Execute Trap capabilities. Older processors rely on an emulation of these features, called Restricted User Mode, and will have a bigger impact on performance.”
That is a real acknowledgement of a performance cost, and it is scoped to older processors that lack the relevant hardware feature and fall back to emulation. Microsoft publishes no figure, and no statement at all about frame rates or games. The percentages circulating in coverage of this are third-party measurements and press framing, not Microsoft statements, and the distinction matters if you are deciding whether to turn off a security feature.
Game Mode. Worth calibrating, because what Microsoft claims for it is narrower than the reputation. Microsoft’s own description: “Game Mode’s performance gain is found in providing, on average, more sustainable framerates for gamers, particularly when your hardware is experiencing resource contention.” Sustainable frame rates under contention. Not higher peaks, and Microsoft has never published a peak frame rate claim for it. That verifiable description is also from 2017 and written about Windows 10, which is worth saying out loud given how much current advice rests on it.
Shape six: only on battery, or only sometimes
Processor power management is the cause people reach for first and it is genuinely the right answer for a narrow set of cases. Microsoft’s description of the mechanism: “Processors change between performance states (P-states) very quickly to match supply to demand, delivering performance where necessary and saving energy when possible.” The two settings that bound this are expressed “as a percentage of maximum processor frequency, with a value in the range 0 – 100”, and the high performance profile is documented as one where “processors are always locked at the highest performance state (including ‘turbo’ frequencies). All cores are unparked.”
Two caveats before you change anything, both from Microsoft. The first is the one that explains why this fix so often does nothing: “If the operating system does not have control over the power management, changing the power plans in Windows will not affect system power and performance.” If your firmware owns power management, Windows settings are decoration. The second: “Capping processor performance at a percentage of maximum requires processor support.”
Note also that this is Microsoft’s server performance tuning documentation. The settings are the same on a desktop, but it is not gaming guidance and we are not presenting it as such.
What to check first, if you want an order
- Put the right numbers on screen. Frame rate alone cannot distinguish any of the six causes above. Temperature, clock speed and video memory use can. Start there.
- Note the shape. Gradual and permanent, or instant and once, or correlated with another application. That single observation removes most of the candidates.
- Capture frame times if the problem is a hitch. A counter averages over a sampling window, so a 117 millisecond spike can be invisible in it. A frame-time capture tool shows it as a spike, and 1% lows explains how to read the result.
- Compare against what it should be doing. If you do not know what your hardware ought to produce, a drop and a normal result look the same. The fps calculator gives you the reference figure.
- Only then change settings. If the cause is thermal, no graphics setting fixes it, and you will have spent an evening proving that.
One thing not to do: turn off a security feature on the strength of a reported percentage. The only thing the vendor has actually published about that cost is a sentence about older processors and emulation, with no figure attached.