GPU Performance Slowly Drops During a Gaming Session, Then Recovers After a Restart? Heat Soak, VRAM Leak & Driver State Diagnosis Guide

Key Takeaways
A graphics card can run normally at the start of a gaming session, then lose frame rate slowly, only to feel normal again after a restart. The cause is something that builds up over time — heat soaking into the case, video memory filling up, or driver state accumulating — and this guide shows how to tell the patterns apart.
The Shape of the Symptom
The symptom this guide covers has a specific shape: frame rate is normal at the start of a session, then falls slowly within the same session, and the fall grows larger the longer the session goes on. A restart brings the performance back, for a while, and then the same slow fall repeats. That shape is the whole diagnosis. It points away from a card that is simply not powerful enough — a card that lacks the power would be slow from the start — and towards something that accumulates over time.
Three accumulating mechanisms can produce this shape, plus a fourth that sits outside the graphics card. Heat soaks slowly into the case, the video memory fills up and never gets released, driver and system state accumulates, and background programs accumulate their own usage. Each one leaves a different trace in the readings, and the readings have to be taken over time, not at a single moment.
Step 1: Confirm the Shape Before Anything Else
Before any testing, confirm that the symptom really matches the shape above. If the frame rate is already low at the start of the session, the problem is not a slow decline — it is a capacity problem from the outset, and the video-memory-capacity guide (#281) or the benchmark-layer guide (#203) is the right path. If the frame rate only falls in certain scenes, that is a scene-load problem rather than a time problem, and the shader-compilation guide (#161) or the menu-load guide (#323) is the right follow-up. If the slow fall within one session is confirmed, continue with this guide.
Step 2: Keep a Time Log, Not a Feeling
Do not judge the decline by feel. Pick one game and one scene, and every ten minutes record the same six readings: frame rate, graphics card temperature, clock speed, power draw, video memory usage, and CPU usage. Keep the log running from the point where performance is normal until the point where the decline is obvious. A log that covers the whole journey from normal to clearly worse is what makes the next step possible; a single reading at the worst moment cannot show a curve, and the curve is the evidence.
Step 3: Read the Curve — Four Patterns
The log will fall into one of four patterns. Count them carefully, because each pattern leads to a different branch, and mixing them up sends the diagnosis down the wrong path.
Pattern A: temperature climbs, clock speed falls
Temperature rises steadily through the session while the clock speed falls with it. This is the heat-related pattern: heat soak, cooling, or case airflow. Go to the heat-soak branch.
Pattern B: temperature and clock speed normal, video memory usage climbs
Temperature and clock speed stay normal, but video memory usage rises steadily through the session. This is the video-memory accumulation pattern: a leak or a capacity build-up. Go to the video-memory branch.
Pattern C: all four readings normal, frame rate falls anyway
Temperature, clock speed, power draw and video memory usage all look normal, yet the frame rate still falls. This is the driver-state, background-accumulation or CPU-layer pattern. Go to the driver and background branch.
Pattern D: temperature, clock speed and power draw fall together, CPU usage stays high
All three graphics readings fall together while CPU usage stays high. This is a CPU or system-layer pattern, not a graphics-card problem. It is named here so it can be recognised and set aside; it is not expanded in this guide.
| Pattern | Temperature | Clock speed | Video memory usage | Frame rate | Direction |
|---|---|---|---|---|---|
| A — heat-related | Rises through the session | Falls as temperature rises | Steady | Falls slowly | Heat-soak branch |
| B — video memory accumulation | Normal | Normal | Rises steadily towards the capacity limit | Falls slowly | Video-memory branch |
| C — driver state or background accumulation | Normal | Normal | Normal | Falls slowly | Driver and background branch |
| D — CPU or system layer | Falls | Falls | Normal | Falls slowly | CPU or system layer, not the graphics card |
Step 4: The Decisive Test — Restart While Still Hot
When the symptom is present, restart immediately, without waiting for the machine to cool down, then go back into the same game and the same scene.
This is the most important test in the whole guide, and it has to be done exactly this way. The restart clears state — video memory usage, driver state, background programs — but it does not clear heat. So the result separates the two:
- Performance comes back immediately, while the machine is still hot. The cause is a state problem: a video memory leak, driver state, or background accumulation. Heat is not the main cause, because the machine is still hot and the performance is back.
- Performance is still slow after the restart, and only improves once the whole machine has cooled down. The cause is heat soak or cooling. State is not the main cause, because the restart cleared the state and nothing improved.
Do not soften this test by waiting for the machine to cool first. Waiting destroys the evidence, because a cooled machine could be better for either reason, and the test would no longer separate them.
Step 5: The Heat-Soak Branch (Pattern A, or the Restart Test Points to Heat)
Heat soak is a real mechanism, and it is not the same thing as simply "the temperature is high". The graphics card's own temperature can reach a steady value quickly, while the air inside the case, the heatsink body, the motherboard and the area around the power supply warm up slowly. At the start of a session the case holds room-temperature air and cooling has room to spare; after a while the whole inside of the case is warmer, the card is already breathing warm air, cooling efficiency falls, and the clock speed is gradually pushed down. The key feature is that the performance fall tracks the temperature inside the case, not the card's own core temperature at any single moment.
The comparisons for this branch:
- Run the same session with the side panel open. This is a test only — do not leave the side panel open as a long-term solution, because it disturbs the case airflow and lets dust build up faster. If opening the panel clearly improves the decline, the case airflow is implicated.
- Clean dust and check the airflow path. Dust and a blocked intake are known causes of throttling, and the cleaning guide (#182) covers the method; it is only pointed to here, not repeated.
- Compare the temperature inside the case over the session, not just the card's own reading, and watch whether the fan speed rises with temperature. The fan-speed guide (#336) covers that check.
- Compare against the card's own start-of-session readings. No universal temperature threshold can be used as a standard here; the only valid comparison is against what the same card showed at the start of the same session, under the same conditions.
Step 6: The Video-Memory Branch (Pattern B, or the Restart Test Points to State)
Video memory filling up and video memory leaking are two different things, and confusing them leads to the wrong fix. Filling up — running out of capacity — means the game never had enough from the start: the symptoms are there from the beginning, and lowering the texture setting improves things immediately. Leaking means the game had enough at the start, but a program — usually the game itself or its driver layer — fails to release video memory it no longer needs, so usage climbs steadily through the session until it approaches the capacity limit and the game starts spilling data into system memory over a much slower path, producing stutter and falling frame rate. Lowering the texture setting can delay the point where the limit is reached, but it does not stop the climb.
The test that separates them is the shape of the video memory usage curve: a flat line at a high level means capacity; a rising slope means a leak.
The comparisons for this branch:
- Lower the texture setting and repeat the same session. Capacity problems improve immediately; a leak only arrives at the limit later.
- Quit the game and start it again, without restarting the whole computer. If performance comes back, that points towards the game's or the driver layer's video memory management.
- Play a different game for the same length of time. If the other game climbs the same way, driver-layer suspicion rises. If only one game does it, that game — or its update — is the strongest suspect (see the game-specific guide, #341).
- A public bug report as an example. A public bug report on the Bohemia Feedback Tracker (DZG-687) records a DayZ case on an Intel Arc B580 with 12GB of video memory: usage started at about 4.7GB with normal performance, and after 30–60 minutes it had risen to about 11.5GB with clearly worse performance, while temperature and a clock speed of about 2850MHz stayed normal — so the cause was neither overheating nor clock reduction. Quitting and reopening the game brought usage back to about 4–5GB and performance recovered immediately. This is cited only as a public bug report, not as an in-house case, and it shows the leak pattern clearly: normal temperature, normal clocks, and usage climbing steadily over time.
Step 7: The Driver and Background Branch (Pattern C)
When all four readings look normal and the frame rate still falls, the fault sits in accumulated state. Over a long session, driver and system state can accumulate — the shader cache grows, driver-internal resources build up, background programs accumulate usage — and performance falls slowly. A restart clears that state, which is why the restart test in Step 4 points here.
The comparisons for this branch:
- Close background programs one at a time, especially a browser with hardware acceleration, recording software, and overlay software. Background accumulation is a real fourth mechanism: a browser left open playing video, for example, can make the same session grow slower over time (see #240). Closing background programs is a necessary comparison, not an optional extra.
- Check the driver version and whether a recent update happened. A clean reinstall with DDU is a last resort, not a first step; the DDU process is covered in the driver guides (#102, #256) and is only pointed to here, not repeated.
- Return any overclock or XMP setting to its default first and repeat the comparison. The overclock-stability guide (#126) covers that step. Do not try to fix a falling frame rate by overclocking further or raising voltage.
Step 8: The Closing Judgement
There are four possible endings, and each one names its next step. None of them is "restart and it is fixed", because a restart only clears state — it does not clear a hardware problem such as dust build-up or ageing cooling, and if every session has to end in a restart, the underlying cause has not been dealt with.
- Heat soak — the Pattern A curve and the restart-while-hot test both point here. The next step is the airflow and cooling checks in Step 5.
- Video memory leak — the Pattern B curve, with usage climbing steadily towards the limit, points here. The next step is the game-versus-driver comparisons in Step 6.
- Driver state — the Pattern C curve, with all readings normal and the restart test pointing to state, points here. The next step is the driver and background comparisons in Step 7.
- Background accumulation — closing background programs restores the decline, which points here. The next step is to identify which program accumulates usage and keep it closed during play.
Boundary Cases
- Slow only at the same time every day, not after a certain length of play. That is a clock-triggered problem — a schedule or a time-of-day temperature pattern — and the same-time guide (#340) covers it. This guide covers time spent in one session, not the time of day.
- Clock speed does not return to idle after gaming ends. That is a leftover state after the session, covered in the post-gaming guide (#326). This guide covers the fall during the session.
- Stutter in the first few minutes, then smooth afterwards. That is shader compilation (#161), and its direction is the opposite of this guide's: rough at the start, smooth later.
- Frame rate falls only in crowded or large scenes. That is scene load, not time accumulation, and it has to be separated from this guide's pattern.
- Laptops. On a laptop, plugged-in versus battery power is the first comparison, because battery power limits power draw. This guide's main subject is desktop machines, and the laptop guide (#150) covers the laptop path.
- Multiple monitors with video playing on the second screen. That is a mixed load, and the multi-monitor guide (#240) covers it.
- Only in summer, never in winter. Room temperature is an amplifier, not the underlying cause. A warmer room changes how long it takes before the fall appears; it does not change which pattern the fall belongs to.
Common Mistakes
- "Falling frame rate means the card is not powerful enough." Wrong. A card that is powerful enough at the start and grows slower over time is not short of power — it has an accumulating problem. A more powerful card could show the same fall.
- "High temperature means thermal throttling." Wrong. Temperature has to be read together with clock speed. If the temperature is high but the clock speed is normal, the card is not being limited by temperature, and no throttling conclusion should be drawn from the temperature alone.
- "A restart fixed it." Wrong. A restart only clears state. The underlying cause — a leak, cooling, or driver state — has not been dealt with, and the same session will fall again next time.
- "Lowering the graphics settings will fix it." Wrong. Lowering settings helps a capacity problem immediately, only delays a leak, and only eases heat soak. It is not a universal fix.
- "High video memory usage means a leak." Wrong. High usage can be normal — a game may deliberately preload textures. What matters is whether the usage climbs steadily over time and approaches the limit, not any single reading.
- "Changing the driver will fix it." Wrong. Changing the driver only helps if the fault is in the driver layer. It does nothing for heat soak or for a leak inside the game itself.
- "Leaving the side panel open is a cooling solution." Wrong. Opening the side panel is a test, not a solution. Left open long-term it disturbs the case airflow and lets dust build up faster.
Frequently Asked Questions
My frame rate is fine at first, then falls slowly. Is my graphics card failing?
Not necessarily. That shape points to something that accumulates over time — heat soaking into the case, video memory filling up, or driver state building up. Keep a time log of frame rate, temperature, clock speed, power draw, video memory usage and CPU usage, and the curve will show which of the four patterns it is.
How do I tell a video memory leak from simply running out of video memory?
Look at the shape of the usage curve. Running out of capacity means the symptoms are there from the start and lowering the texture setting improves things immediately. A leak means usage climbs steadily through the session from a normal start, and lowering the texture setting only delays the point where the limit is reached.
Why must the restart test be done while the machine is still hot?
Because the restart clears state but not heat. If performance comes back immediately while the machine is still hot, the cause is a state problem. If it stays slow until the machine has cooled down, the cause is heat. Waiting for the machine to cool first destroys that separation.
My temperature looks high. Does that prove thermal throttling?
No. Temperature has to be read together with clock speed. If the clock speed is normal while the temperature is high, the card is not being limited by temperature. Read all four readings — temperature, clock speed, power draw and video memory usage — together, because any one of them on its own can mislead.
High video memory usage — is that always a leak?
No. A game may deliberately preload textures, so high usage on its own is normal. A leak is shown by usage climbing steadily over time towards the capacity limit, not by any single reading. The public DayZ bug report (Bohemia Feedback Tracker DZG-687) is an example of that climbing pattern.
Lowering my graphics settings helped a little. Does that mean the problem is fixed?
Not necessarily. Lowering settings helps a capacity problem immediately, only delays a leak, and only eases heat soak. If the slow fall still appears, just later in the session, the underlying cause is still there and the pattern in the time log will show which one it is.
Should I leave the side panel open to keep the card cool?
No. Opening the side panel is a useful test for case airflow, but it is not a solution. Left open long-term it disturbs the case airflow and lets dust build up faster.
Getting Help
A frame rate that falls slowly within one session and recovers after a restart is a pattern problem, and the pattern has to be read from a time log, not from a feeling. Four readings taken every ten minutes will show which of the four patterns it is, and the restart-while-hot test will separate state from heat. When the fault points to the hardware layer — cooling that no longer holds up, or a card whose behaviour needs hands-on inspection — the card should be brought in for inspection. HK GPU Fix is at Room 1209, Dali Plaza, Shek Mun. For enquiries, call 2186 8010 or WhatsApp 6558 6806. There is no on-site service, so the card should be brought to the shop for inspection.