By Walter Jobs, Technology writer and editor · Published 9 October 2026
Short answer: a frame requires two jobs done one after the other by two different chips, and they finish at different times. Whichever finishes last sets your frame rate. That is the whole mechanism, and it means every machine has a bottleneck at every moment. The useful questions are which part it is and how large the gap is, not whether one exists.
This page is about why it happens. If you want to know whether it is happening to you, 7 signs of a CPU bottleneck is the diagnostic version.
How a frame is actually built
Every frame goes through two stages that cannot overlap for the same frame.
Stage one, the processor. It reads your input, advances the physics, runs the AI, updates every object in the world, decides what is visible from where you are standing, and issues a list of draw calls describing what should be rendered. This is bookkeeping about a world, and its cost depends on how much is in that world.
Stage two, the graphics card. It takes that list and fills in pixels: geometry, textures, lighting, shadows, reflections, post processing. Its cost depends on how many pixels there are and how expensive each one is to compute.
The card cannot start on a frame the processor has not described yet, and the processor cannot skip ahead indefinitely. So the frame takes as long as the slower stage, and your frame rate is set by whichever chip is lagging.
Why they never match
Here is the part that explains everything downstream. The two stages respond to completely different variables.
| Change this | Processor cost | Graphics card cost |
|---|---|---|
| Raise resolution from 1080p to 4K | Essentially unchanged | Roughly four times higher |
| Add a hundred units to a battle | Much higher | Slightly higher |
| Turn shadows and reflections up | Barely moves | Much higher |
| Increase draw distance | Higher | Higher |
| Raise the frame rate target | Proportionally higher | Proportionally higher |
Because the two columns respond differently, no pair of parts can be matched for all situations at once. A combination that balances perfectly at 1440p in a shooter will be processor limited at 1080p in a strategy game and card limited at 4K. The balance is a property of the situation, not only of the hardware.
This is why a single bottleneck percentage attached to a pair of parts, with no resolution and no game attached, is not telling you much. That argument is made at length in does a bottleneck percentage really matter.
Why the processor side cannot simply be made parallel
The obvious objection is that processors have had many cores for years, so why does the processor stage not just scale?
Because a large part of that work is inherently sequential. The physics result depends on the AI decision, which depends on your input, which depends on the previous frame’s state. You cannot compute step four before step three. Game engines do push what they can onto other threads, including audio, asset streaming, animation and particle work, but the main game thread remains a chain of dependent operations that runs on one core.
That is why a sixteen core processor can be the limit while showing 25% total usage, and why single core speed and cache size predict gaming performance far better than core count does. Beyond about eight cores, extra ones have very little to contribute during gameplay.
The graphics stage has the opposite character. Filling two million pixels is two million independent problems, which is exactly what thousands of small parallel units are for. This asymmetry in how the two jobs divide is the deepest reason the two halves scale differently.
The things that actually create a gap
- Buying the two parts years apart. The most common cause by far. A new card dropped into a four year old machine inherits that machine’s processor ceiling.
- Splitting a budget unevenly. Spending heavily on the card and minimally on the chip produces a card that cannot be fed at the resolutions that card was bought for.
- Playing at low resolution on strong hardware. Nothing is wrong with the parts; the settings are asking the processor for everything and the card for very little.
- Chasing a high refresh rate. Doubling your target frame rate halves the time the processor has per frame, and that budget is the real constraint at 144 Hz and above.
- Genre. A strategy game simulating thousands of entities asks for a different machine than a corridor shooter does, with no change in resolution.
- Starving the processor by accident. Single channel memory, a memory profile left switched off, or a background task taking the core the game needs, all produce a processor limit on hardware that should not have one.
Why a zero percent bottleneck is not a goal
Perfect balance would mean both chips finishing at precisely the same instant in every scene, at every resolution, in every game. Since the two respond to different variables, that is not achievable even in principle, and chasing it wastes money.
What a well built machine looks like is a card that is the limit most of the time, with a processor that has enough headroom not to become the limit in the heaviest scenes at your resolution. The card being the slower part is the correct outcome, for the reasons set out in CPU or GPU bottleneck, which one is worse.
To see where a specific pair falls, put both parts into the bottleneck calculator, which computes each chip’s ceiling separately at each resolution rather than applying a fixed weighting.
Frequently asked questions
Does every PC have a bottleneck?
Yes, necessarily. Something has to be the slowest part of a sequential process, and that something sets your frame rate. The question is never whether you have one but whether the gap is large enough to be worth money, and whether the limiting part is the one you would choose.
Why does more cores not fix a CPU bottleneck?
Because the critical path through a game’s frame is a chain of dependent steps that must run in order on one core. Engines offload what they can, but the main thread stays serial. This is why gaming performance tracks single core speed and cache far more closely than core count, and why a chip with fewer, faster cores often beats one with more, slower ones in games.
Can software updates change which part is my bottleneck?
Yes, in both directions. A graphics driver can reduce the processor overhead of submitting work, and a game patch adding simulation detail increases it. A newer engine version can move the balance noticeably on identical hardware, which is why a machine that was card limited at launch can be processor limited two years later.
Is a bottleneck the same as a system being slow?
No. A bottleneck describes which component sets the ceiling, not how high the ceiling is. A very fast machine has a bottleneck too, it just has it at 200 frames per second. Describing a machine as bottlenecked says nothing about whether its performance is acceptable.
Why do two people with the same parts report different bottlenecks?
Because the balance depends on what is being asked. Different resolutions, different games, different settings and different memory configurations all move which chip finishes last. Two identical machines genuinely can be limited by different components, and both reports can be accurate.
The relationships described here are how the rendering pipeline works rather than measurements. Where this article refers to specific figures, they come from our own model, set out on how we calculate this.