GPU Slower After a Manual Driver Update, FPS Down and Clock Stuck Low: Driver Rollback, Settings Reset, Shader Cache and Power Plan Diagnosis Guide

Key Takeaway
After a manual driver update, the same game in the same scene runs slower with the clock stuck low. Run one rollback comparison first: reinstall the previous driver version, retest the same sixty seconds, and see whether the frame rate returns. That single result separates a driver-version fault from reset settings, a rebuilding shader cache, or a changed power plan.
The Shape of the Problem: Slower Only After a Manual Update
This piece deals with one tight pattern only, and the whole diagnosis stays inside it. A game was running as it should. Its owner fetched the newest driver from the vendor's site or the vendor's own app and put it in by hand. From that moment, the same title in the same spot runs plainly slower, the frame rate has dropped, and the card's clock no longer climbs the way it once did. The desktop still loads. No blue screen shows up, nothing restarts on a loop, and the machine still boots. What changed is performance, and the change starts exactly at that hand-installed driver.
Three conditions lock the shape in place, and every one of them has to be true. First, the trigger was a hand-installed newer vendor driver, put in by the owner. Windows Update did not push it onto the machine, and nothing was swapped behind the owner's back. Second, the before-and-after test runs on one machine: one title, one spot, one set of settings, slower now than before the update. Third, the machine is steady in every other way. It reaches the desktop and keeps running, and slower games with a clock that stays low are the only fresh symptom.
Drawing that line matters, because the usual mistake with this shape is to treat "the driver was updated" as a finished answer. It is not finished. Putting in a driver swaps more than one thing at once. It swaps the driver files, and it can also rewrite control-panel choices, reset profiles kept per game, alter the power preference, reset the display refresh rate, and wipe the stored shader cache. Any one of those swaps can slow a game. So the question here is not "did the update cause it" in general. It is "which item the update touched is the one slowing the game" — and the answer may sit in the driver release itself, in a choice the installer reset, in a shader cache that must be rebuilt, or in a power plan that moved. Each answer calls for its own test, and the rollback test is the one that pulls the first answer apart from the rest.
A crash in speed with no update behind it belongs somewhere else entirely. A game that was fine, then begins hitching without warning, or a frame-rate figure that slides away midway through a session, is trouble growing across time, and trouble growing across time belongs to another set of causes. What this guide follows is the mirror image: performance that stepped down at one identifiable instant — the moment a newer driver was fitted by hand — and has stayed down ever since. The checks below weigh two settled states against each other — the fresh release set against the older one, the first outing set against the third, one option switched on set against the same option switched off — because a fault pinned to a single event is found by shifting a lone factor and seeing whether the settled state shifts back.
A further boundary belongs up front, before any check runs. Choosing a driver release in general is not taught here. One move is used from that topic — putting the immediately previous release back in place a single time, as a test — and nothing further. Shader compilation on its own is not taught here either. One move is used — playing one spot three times and watching whether the stutter fades — and nothing further. Windows power plans on their own are not taught here either. One move is used — looking at the three power positions and altering one of them a single time, as a test — and nothing further. The piece stays inside its own pattern by using one move from each nearby topic and leaving those topics themselves untouched.
Why a Newer Driver Can Run an Old Game More Slowly
A driver update is not one file being exchanged for a quicker one. It is a whole bundle of files being replaced together: the core driver, the control panel, the profiles kept per game, the way the shader cache is handled, and the power-management policy. A fresh release is usually tuned for newer titles and newer features, and while doing that work it can make an older title, an older engine, or one particular hardware pairing run more slowly. Vendors admit this themselves: their release notes carry a Known Issues or Open Issues section, and a regression in one title, one API, or one pairing is exactly the sort of entry that turns up there. A regression is therefore a genuine mechanism, not an excuse dreamed up by the owner, and it is also not a judgement on the fresh driver as a whole. A new release can run one title faster and another title slower, and the only evidence that carries weight is a before-and-after test on one machine, in one title, in one spot.
The installer rewrites choices as well, and that second mechanism explains the cases where the release itself is innocent. An install can send the control-panel power preference from a maximum-performance choice back to its adaptive default. It can reset the shader-cache choice. It can overwrite profiles kept per game, so a title's quality choices, refresh rate, or resolution are no longer what the owner set. A clean install — the installer's clean-install choice, or a factory reset on some vendors — clears even more thoroughly, taking the old choices away with the old files. The result is that the release and the choices are two separate variables, and they must be tested apart from each other. Altering both together and then deciding that "the rollback failed" proves nothing, because the test cannot tell which of the two alterations mattered.
The third mechanism is the shader cache. Before a title can paint a picture, its shaders must be turned into machine code the card understands, and the finished results are kept in a cache so they can be used again instead of being compiled afresh. Putting in a new driver wipes the old cache — NVIDIA says this outright — so the first time a title is entered after the update, and the first time a fresh spot or a fresh effect appears, the system must compile those shaders again. While the cache is being rebuilt, frame times can leap and the title can stutter. That is a rebuilding process to expect, not a lasting loss of speed. The check that pulls it apart from a genuine regression is repetition: play the same spot three times, and watch whether the stutter fades. If it fades, the cache was being rebuilt. If it stays just as bad on the third run as on the first, the cache is not the explanation.
Power planning forms the fourth mechanism, and its three settings have to be viewed as a set. An update can move the Windows power plan, the energy choice within the vendor panel, and, on a portable machine, how it behaves on wall power versus on its battery. Examining just one of the three can leave the other two unnoticed. On a balanced plan, or under an energy-saving preference, the system sheds speed more readily to spare power, and under a long load the clock may climb more tardily and level off lower. This layer stands next to the guide about a driver swapped in by Windows Update, which covers power plans in full; this guide takes only the three-setting check from it, and does not repeat that guide's power-plan diagnosis.
One more mechanism must be guarded against while the checks run, even though it is not the topic of this piece. After going back to an earlier driver release, Windows Update can put a newer release back in place on its own, sometimes within hours, and that would quietly undo the comparison. The release number must therefore be noted down before the rollback test and noted down again after it. If the two numbers differ, the comparison is void, because the driver under test is no longer the one the owner meant to test. This is a guard on the test, not the topic of the piece: the guide for a driver replaced by Windows Update treats that mechanism in full, and this piece uses only the release-number check.
Five Deciding Comparisons That Narrow the Field
Five tests carry the diagnosis here, and each test has two possible results. They must be run in order: every result shrinks what the next test has to pull apart, and the opening test — the rollback test — carries the most weight of any check here. It does not rest on the absolute size of any one reading. It rests only on what moves when one variable moves, in one stretch, on one machine.
The first test is the rollback comparison, and it is the main check of this piece. The current driver release number is noted down. The previous release — the one in place before the hand update — is put back, either through the roll-back choice in Device Manager or by fetching the older release from the vendor's site and installing it cleanly. The same title, the same spot, and the same sixty seconds are then tested again. If the frame rate comes back to its level before the update straight away, the driver-release layer is the answer: the new release itself regressed, or the new release does not suit this title and this pairing. If the title is still slow on the old release, the release itself is not the culprit, and the diagnosis moves to the choices, the cache, the power plan, and the hardware layers. The power of this test is that it alters one variable only — the driver release — and leaves everything else untouched. A rollback used this way is a test, not advice to stay on the old release for good. A newer release may gain a hotfix later, and the release notes are the place to look for one; this piece states that as something to check, not as a promise with a date, because no vendor has promised one.
The second test sets the first outing against the later, settled outings. The hitching appears only when a new stretch is entered for the first time, or a new effect is seen for the first time, after the update, and it fades when that same place is revisited. That shape belongs to the shader-cache rebuilding layer. The check is to replay that identical stretch a trio of times and jot the frame-rate figure after every pass: a figure that improves pass by pass means the cache is still being assembled, and any verdict has to wait until the assembly is done; the same sluggishness on every one of the three passes rules the cache out, and the hunt moves on. A first-outing hitch must never be written up as a permanent loss of speed, and emptying a cache must never be written up as a speed gain: emptying only throws the stored work away, and the price is that the next outing has to compile it again.
The third test is the rollback that did not help. The old release is back, and the title is still slow. That result points at something the update altered besides the release: the control-panel power preference was reset, the in-game quality choices were reset, the Windows power plan was altered, or the display refresh rate was reset. Each of those choices is checked against what it was before the update, where that was recorded, or put back to its default one by one, with a fresh test after every single change. This layer is the one most often misread as "the old release is broken too". It is not. The release went back, but the choices did not, and the choices are a separate variable that must be put back separately.
The fourth test is one title against all titles. Only one title is slow, and the rest run normally: the fault sits in the title or engine layer, either because the new driver does not suit some function of that one title, or because the title itself needs an update. Every title is slow: the fault sits in the driver or system layer. This test decides which road the diagnosis takes next, and it cannot be skipped, because the two roads lead to different checks.
The fifth test weighs clock, power draw, and temperature against each other. The clock sits low, the power draw sits low, and the temperature sits in its normal band. That combination can be read two ways, and the test separates the readings by watching whether they shift in step as the load shifts. If the clock rises as the load rises, the card is simply idle — the title is asking little of it, or a frame cap is holding it back — and a low clock is an unremarkable response. If the clock refuses to budge however heavy the load grows, the card is being pinned down, and the energy preference or the driver's power state becomes the suspect. A lone clock figure cannot separate the two readings.
| Test | What is compared | Result pointing at the release or the cache | Result pointing at the choices, power, or hardware |
|---|---|---|---|
| Rollback comparison | Same spot, previous driver release put back in place | Frame rate comes back to its level before the update straight away | Still slow on the old release; the release is not the culprit |
| Opening run against settled run | Same spot played three times | Frame rate gets better each time; stutter fades | Same slowness on all three runs; not the cache |
| Rollback did not help | Choices checked one by one after the rollback | Not applicable; the release is already excluded | Frame rate comes back after one choice is put back; that choice was the layer |
| One title against all titles | Other titles tested the same way | Only one title slow; title or engine layer | Every title slow; driver or system layer |
| Clock, power draw, temperature | Three figures followed in step as the load shifts | Clock rising as the load rises; the card is simply idle | Clock refusing to budge however heavy the load becomes; pinned down |
Why the Opening Run Stutters and Why One Clock Reading Proves Nothing
The shader cache explains why a game can seem slower right after an update before sorting itself out, and getting this wrong is among the commonest errors in this diagnosis. The cache stores compiled shaders so a stretch need not be compiled anew on every visit. A fresh driver clears the old cache, so the first visit to a title, and the first visit to each new stretch, has to compile those shaders anew. While that compiling is going on, frame times jump and the title hitches. The hitching is genuine, and it is also brief: once the cache is assembled, that same stretch runs smoothly again. The check is repetition in one stretch, three passes, with the frame-rate figure jotted after every pass. A steady climb across the three passes is the cache being assembled. No climb across the three passes is not the cache, and the hunt must move on. Two errors have to be steered clear of here. The first is to call a first-outing hitch a permanent regression: it is a rebuilding process, and judging the fresh driver on its first outing alone is judging it before any cache has been built for it. The second is to present emptying the cache as a way to make a title quicker: emptying only deletes the stored work, and the next outing pays for it by compiling again.
The clock figure deserves equal caution, for a separate reason. A card's clock climbs and drops with the load laid on it. When the card is idle — because the title is light, because a frame cap is holding it, or because the processor is not supplying work quickly enough — a low clock is an unremarkable response, not a fault. The boost figure listed in the specification sheet is a peak the card can touch when the load asks for it. It is not a speed the card is pledged to sustain at every moment. Sizing up a clock speed by itself, without the usage, wattage, and temperature figures alongside it, is therefore an error. A low clock beside low usage, low power draw, and an ordinary temperature can describe a card sitting idle, and it can also describe a card pinned down by a power state that has not lifted — and the two shapes look identical in one snapshot. The gap only appears over time, as the load shifts. An idle card raises its clock as the load climbs. A pinned-down card does not raise its clock as the load climbs. That is why the fifth test follows the figures shifting in step instead of reading one figure a single time. One clock figure cannot separate the two, and one usage figure cannot either.
One more link deserves attention: how a reset option and a power state act on each other. When the control-panel energy preference is reset from a top-performance option to its energy-saving default, the card may keep a lower clock under load, and that lower clock then seems like a second fault, when in truth it is a knock-on effect of the reset. That is why the power-state test belongs after the roll-back test and the options check, not ahead of them. If the roll-back already shows the frame rate returning on the older release, the low clock is accounted for by the release layer, and changing energy options will add nothing. Running the power-state test ahead of the others, and treating the low clock as the culprit, steers the hunt towards a layer that is merely obeying the true limit.
The Five Suspicion Layers
Every instance of gaming slowing after a hand-fitted driver update lands in one of the five layers set out below. The layers appear in the sequence the evidence most often points towards, but that sequence settles nothing: a layer is named only once a test has backed it or cleared it. The sources do not agree on which layer shows up most often, so this guide places none of them first. Each layer arrives carrying the evidence that would point towards it, and the evidence decides.
| Layer | What it means | Evidence that points at it |
|---|---|---|
| First layer: driver-release layer | The new release itself regressed for this title, this API, or this pairing | Rollback test: the old release runs normally straight away and the new release is slow again; the vendor release notes list the same symptom under Known Issues or Open Issues |
| Second layer: choices-reset layer | Control-panel power preference, per-title profile, or display refresh rate and resolution reset by the installer | Still slow after the rollback; the frame rate comes back after one choice is put back, one at a time |
| Third layer: shader-cache layer | The new driver wiped the old cache, and the opening runs must compile shaders again | Only the first entry into a fresh spot stutters; the same spot played three times gets better each time |
| Fourth layer: power-plan layer | Windows power plan, control-panel energy preference, or portable battery mode moved by the update | Clock refusing to budge however heavy the load becomes; changing the energy option a single time as a test lets the clock track the load |
| Fifth layer: coincidental hardware or other layer | Something else that happened on the same day: dust reaching a limit, the title itself updated, a background program taking resources, a temperature limit | Still slow after the rollback and the choices restoration; temperature or fan readings abnormal; slow across titles |
The first layer is the driver-release layer, and it is the layer the rollback test is built to find. A new release can regress for one title, one API, or one pairing, while running other titles perfectly well. The evidence is the rollback test itself: the old release runs normally straight away, and the new release is slow again. Supporting evidence comes from the vendor's release notes, where the same symptom may be listed under Known Issues or Open Issues. The follow-up is to stay on the old release while watching for a hotfix — not as a lasting solution, but as a holding position. One published guide states that hotfixes usually arrive within one to three weeks; that is that site's claim, not a vendor promise, and this piece does not repeat it as a date. The release notes are the place to watch, and no date can be stated here.
The second layer is the choices-reset layer. The installer rewrites control-panel choices: the power-management preference can return from maximum performance to its adaptive default, the shader-cache choice can be reset, profiles kept per title can be overwritten, and the display refresh rate or resolution can be altered. A clean install clears even more thoroughly. The evidence is that the title stays slow after the rollback, and the frame rate comes back after one choice is put back — one choice at a time, with a fresh test after each change, so that the test can tell which choice mattered. This layer sits beside the guide for a driver replaced by Windows Update, which treats settings resets in full; this piece uses only the one-at-a-time choices check, and does not redo that guide.
The third layer is the shader-cache layer. The new driver wipes the old cache, and the opening runs after the update must compile shaders again. The evidence is narrow and specific: only the first entry into a fresh spot stutters, and the same spot played three times gets better each time. This layer sits beside the guides for shader-compilation stutter, which treat that topic in full; this piece uses only the cache-rebuilding mechanism and the three-run check, and does not redo their diagnosis.
Fourth in the list comes the power-plan layer. Windows power planning, the panel's energy preference, and a portable's battery mode can all be moved by the update, and all three have to be examined as a set, because examining just one can leave the other two unnoticed. The evidence is a clock that refuses to budge however heavy the load becomes, plus one trial in which changing the energy option a single time lets the clock track the load. Every change in this layer is a trial, not a fix: the option is changed once, the outcome is written out, and the option is restored. A change that lets the clock climb shows the trouble sits in the power-plan layer. It does not show the option ought to stay changed.
The fifth layer is the coincidental hardware or other layer, and it is an exclusion layer. Something else may have happened on the same day as the update: dust in the cooling reaching a limit, the title itself receiving an update, a background program taking resources, or a temperature limit being reached. The evidence is that the title stays slow after the rollback and after the choices restoration, with abnormal temperature or fan readings, or with slowness across titles that the release and choices layers cannot explain. This layer must never be written as "the update definitely had nothing to do with it". It can only be written as "the release and the choices have both been excluded, so the range must widen".
The Seven Diagnostic Steps
Go through the seven checks in the order they are printed. Each check gives back an outcome that chooses the check that comes next. Jumping to a driver reinstall, or to a hardware swap, ahead of the first three checks usually squanders the attempt: without a baseline, a look at the release notes, and a first-outing test, a later change settles nothing, even when the frame-rate figure shifts.
| Check | What to do | How to read the result |
|---|---|---|
| Check 1 | Pin down the stretch and record the baseline: one title, one stretch, one sixty-second run; record mean frame rate, clock, power draw, temperature, card usage, frame time, 1% low, and driver release digits | With no pinned stretch, no later test means anything; unlike stretches place unlike loads |
| Check 2 | Note the release number and read the release notes: copy the driver vendor, date, and release from Device Manager; read the vendor release notes for Known Issues or Open Issues | The same symptom listed lifts the suspicion on the release layer; its absence does not exclude the release layer — carry on to the test |
| Check 3 | Run the opening-run test: play the same spot three times and note the frame rate each time | Getting better each time means the shader cache is being rebuilt — play more times and let the cache build before judging; the same slowness on all three runs means it is not the cache — move on |
| Check 4 | Run the rollback test, the deciding check: put the previous driver release back in place, retest the same spot, and note the release number down again afterwards to confirm Windows Update has not quietly put the new release back in place | Frame rate back to its level before the update means the driver-release layer — stop here; still slow means move on |
| Check 5 | Run the choices-restoration test: look at the control-panel power preference, the in-game quality choices, the Windows power plan, and the display refresh rate and resolution, putting back one item at a time and retesting after each one | Whichever item brings the frame rate back names the layer |
| Check 6 | Run the range test: is only one title slow, or are all titles slow | Only one title slow means the title or engine layer — look for a title update and the title's own choices; all titles slow means the driver or system layer — go back to the fourth and fifth layers |
| Check 7 | Halt line: the roll-back, the settings put back, and the cache rebuilt are all finished and the title is still slow, and the temperature or fan figures are off, or picture corruption, a black screen, or a crash has appeared | Carry the records — the release digits on both sides of the test, the three frame-rate figures, the clock, power, and temperature figures, and the checks already tried — for in-store inspection; do not keep trying driver releases one after another, and do not install old drivers found online one by one |
The first check lays the footing for every check that follows, and it happens ahead of any test running. Pin down the stretch: one title, one saved position, one view, one sixty-second run of play. Write out the baseline in full: the driver release digits, the mean frame rate, the clock, wattage, and temperature, the card usage, the frame-time graph, and the 1% low figure where the tool supplies one. Do not set a quiet indoor stretch against a busy outdoor stretch: the two stretches load the processor and the card unlike, so a test between them proves nothing. Where the frame rate from before the update was never written out, say so plainly: memory is not a measurement, and it cannot stand in for a baseline. The baseline is the measure for every later check. Without it, a later figure cannot be judged as a shift or as no shift, because there is nothing to weigh it against.
The second check is the release-number and release-notes look. The driver vendor, date, and release are copied from Device Manager. The vendor's release notes for the new release are then read, looking for the same symptom under Known Issues or Open Issues. If the same symptom is listed, the suspicion on the release layer climbs. If it is not listed, the release layer is not excluded — release notes list what the vendor knows about, not everything that can happen — and the diagnosis carries on to the test. The release number noted down in this check is also the record that the fourth check will be tested against, to confirm that Windows Update has not quietly swapped the driver in between.
The third check is the first-outing test, and it sits ahead of the roll-back for a reason: a cache that is still being assembled can make the roll-back test meaningless, because the first outing on the older release would hitch too. That identical stretch is replayed a trio of times, and the frame-rate figure is jotted after every pass. If the frame-rate figure climbs after every pass, the shader cache is still being assembled, and any verdict has to wait until the assembly is done — replay the stretch more times and weigh the figures again. If the sluggishness is identical on every one of the three passes, the cache is ruled out, and the hunt moves to the fourth check.
The fourth check is the roll-back test, and it is the deciding check of this guide. The earlier driver release — the one in place ahead of the hand update — is fitted back, either through the roll-back option in Device Manager or by fetching the older release from the vendor's site and fitting it cleanly. That identical stretch is tested again across the same sixty seconds. If the frame rate returns to its level from before the update, the driver-release layer is confirmed, and the hunt halts at this check. If the title is still slow, the release itself is not the culprit, and the hunt moves to the fifth check. After the retest, the release digits are written out again, to confirm Windows Update has not quietly fitted the fresh release back in place in the meantime. If the two sets of digits differ, the test is void and has to be run again. A roll-back used this way is a test. It is not advice to remain on the older release for good, and it must not be written as if the older release were always better than the fresh one.
The fifth check is the choices-restoration test. The control-panel power preference, the in-game quality choices, the Windows power plan, and the display refresh rate and resolution are looked at one at a time, each put back once, with a fresh test after every single change. Whichever item brings the frame rate back names the layer. The discipline here is a single variable at a time: putting back several choices at once and then seeing an improvement proves nothing, because the test cannot tell which choice mattered. The choices altered for the tests are put back to their earlier state where the test shows they were not the cause, and the result of every single change is noted down.
The sixth check is the range test. Only one title is slow, or every title is slow. Only one title slow means the title or engine layer: check whether the title itself has an update, and check the title's own choices, because the new driver may not suit some function of that one title. Every title slow means the driver or system layer, and the diagnosis goes back to the fourth and fifth layers. This check is deliberately placed after the rollback and the choices checks, because a range result on its own cannot tell whether the release or the choices are at fault — it can only tell where to look next once those two have been tested.
The seventh check is the halt line. The roll-back, the settings put back, and the cache rebuilt are all finished, and the title is still slow. The temperature or fan figures are off, or picture corruption, a black screen, or a crash has appeared beside the sluggishness. At that point, halt the testing, and carry the records for in-store inspection: the clock, power, and temperature figures, the list of checks already tried, the release digits on both sides of the test, and the frame-rate trio recorded in the third check. Do not keep trying driver releases one after another, and do not install old drivers found online one by one. Trying releases at random adds changes without adding insight, and it leaves the cause still unidentified.
Six Misleading Ideas to Set Aside
Misleading idea one: a newer driver is always faster. This is wrong. A new release can regress for one title, one API, or one pairing, while an older release runs that title more steadily. Release age is not a quality guarantee. The evidence can only come from a before-and-after test on one machine, in one title, in one spot. "The new release is always better" and "the new release is always worse" are both wrong, and both must be replaced by the rollback test.
Misleading idea two: stutter on the opening run after an update means the new driver has lastingly slowed the title down. This is wrong. An opening-run stutter is most often the shader cache being rebuilt, because the new driver wiped the old cache and the first entry into a fresh spot must compile shaders again. Play the same spot three times before judging: if the stutter fades each time, it was the cache. Judging the new driver on its opening run alone, before its cache exists, misjudges the release layer.
Misleading idea three: rolling back is giving up. This is wrong. A rollback is the strongest test in this piece: it proves whether the problem follows the release or not, because it alters one variable only. Staying on the older release while watching for a hotfix is a reasonable holding position, not a failure. The rollback is a test that names the layer. It is not an admission that the machine cannot be fixed.
Misleading idea four: reinstalling the driver is the opening move. This is wrong. First write out the release digits, read the release notes, and run the first-outing test. Reinstalling ahead of the layers being pulled apart wipes out the evidence the tests need — the release digits, the cache state, and the options exactly as the installer left them. The same reasoning applies in the guide for a state that fails to change after gaming: reinstalling ahead of diagnosing destroys the test.
Misleading idea five: a small drop in frame rate proves a regression. This is wrong. One published guide states that a change under around 3% falls within run-to-run variation — that is that site's claim, and it is not presented here as a general standard, because no general standard is agreed across the sources. The judgement must rest on whether the same spot, tested several times, stably reproduces the loss, not on one number. One frame-rate figure, taken a single time, proves nothing on its own.
Misleading idea six: if one clean reinstall with a driver-cleaning tool does not help, clean again. This is wrong. One complete cleaning and one clean reinstall is enough. Repeating the same cleaning process does not alter the result; it only hides the real cause by adding more alterations to undo. The guide for driver cleaning states this directly, and this piece follows it: one complete cleaning, one clean reinstall, and then the result is read.
Results Table: Reading Each Outcome
The table that follows pulls the checks together. Each row opens with something that can be seen — a frame-rate figure, a clock shift, an option changed — and closes with the layer that outcome indicates. It comes after the seven checks, as a cross-check on the diagnosis, without taking their place.
| Outcome observed | Layer indicated | Next move |
|---|---|---|
| Going back to the previous release brings the frame rate back straight away | First layer: driver-release layer | Stay on the older release while watching the release notes for a hotfix; do not treat the old release as a lasting solution |
| Same spot played three times, frame rate getting better each time | Third layer: shader-cache layer | Play the spot more times and let the cache build before judging the driver |
| Same slowness on all three runs of the same spot | Not the cache | Move to the rollback test |
| Still slow after the rollback; one restored choice brings the frame rate back | Second layer: choices-reset layer | The restored choice names the layer; note which one it was |
| Clock refusing to budge however heavy the load becomes | Fourth layer: power-plan layer | Change the energy option a single time as a trial and restore it |
| Clock climbs when the load climbs | Card waiting, not held low | Go back to the layers above; the low clock is an ordinary response |
| Only one title slow, others normal | Title or engine layer | Look for a title update and the title's own choices |
| Every title slow | Driver or system layer | Go back to the fourth and fifth layers |
| Still slow once the roll-back, settings put back, and cache rebuilt are done, with off temperature or fan figures | Fifth layer: coincidental hardware or other layer | Halt the testing and carry the records for inspection |
| Release number altered between the two readings around the rollback | Test void | Windows Update might have put a newer release back in place; run the test again with the numbers checked |
The last two rows deserve a closer read, because they spell out what each test can and cannot establish. A rollback that brings the frame rate back proves that the problem follows the release — but it does not prove that the choices are innocent, because a choices reset can exist at the same time as a release regression. The choices-restoration test must still be run before the two layers can be pulled apart, and this piece states that limit plainly. A release number that altered between the two readings proves nothing about the driver at all: it proves that the test itself was not run on the release the owner meant to test, and the test must be run again with the numbers checked.
Where the Sources Part Ways
The sources behind this guide part company with each other, and the points where they split need plain setting-out rather than smoothing away. The first parting is over whether a roll-back is a repair or a test. Two published guides call a roll-back the quickest repair. The view taken here is different: a roll-back is a test. It shows the trouble follows the release, and it does not advise remaining on the older release for good. The parting matters, because a reader who takes the test for a repair stops the work right where the layer has just been identified, and never looks to see whether a hotfix has since appeared.
The second split is about whether a driver-cleaning tool is necessary. One published guide presents a thorough cleaning as the complete solution. The guide for driver cleaning states that one complete cleaning is enough, and that repeating it does not alter the result. This piece follows that reading: one complete cleaning, one clean reinstall, and then the result is read. Repeating the same process adds alterations without adding information.
The third split is about small frame-rate changes. One published guide states that a change under around 3% falls within run-to-run variation. That is that site's claim only, and no other source consulted here confirms it as a general standard. This piece does not adopt it as a rule. It asks the reader to judge by whether the same spot, tested several times, stably reproduces the loss, rather than by one number against a threshold that the sources do not agree on.
The fourth split is about hotfix timing. One published guide states that hotfixes usually arrive within one to three weeks. That is that site's claim, not a vendor promise, and no vendor has promised a date. This piece does not repeat it as a date. It says only that the vendor may release a hotfix, and that the release notes are the place to watch for one.
The fifth parting is over which cause shows up most often. One source places the release regression first. A second places the shader cache first. A third voice starts by separating a micro-stutter from a full frame-rate fall, treating those as unlike shapes. The three are describing unlike cuts through the same trouble, not three orderings of one list, and no settled view emerges. This guide places none of them first. It sets out five layers, every one backed by its own evidence, and it never singles out one of them as the most frequent cause.
Outside the Scope: Patterns for Other Guides
The table that follows gathers the shapes that look like this one but sit with other guides. Each row names the shape, the guide it sits with, and why it is not this guide. The table belongs to the diagnosis: when the shape before the reader matches a row here more closely than the shape in the opening section does, the reader has the wrong guide and ought to switch to the right one.
| Pattern | The guide that covers it | Why it sits outside this piece |
|---|---|---|
| Driver fitted by Windows Update, driver vendor switched to Microsoft, release stepped back | #189 and #310, Windows Update driver replacement | The trigger is an automatic system update, not a hand fitting by the owner |
| Blue screen, repeated restarting, or failure to reach the desktop right after putting in a driver | #497, driver loading layer | The premise here is reaching the desktop with only slower performance, not a boot failure |
| Only sluggish after sleep, fine again once the game restarts | Guides #284 and #501 on wake-triggered faults | Waking is the trigger here, not a driver fitted by hand |
| Clock or fan speed fails to drop back to idle after gaming, or fails to climb when a game begins | Guide #326 on state-transition residue | Here the fault is a state that will not switch over; that guide expressly keeps post-update performance loss outside its range |
| A sharp hitch, or a big drop in frame rate from a normal level | Guide #136 on sudden stutter plus frame-rate collapse | Here the fault arrives suddenly, unlike a steady loss that began at one fitting |
| Title grows slower the longer it runs | Guide #356 on progressive slowing | Here the fault is a gradual shift over time, not a loss that began at one fitting |
| First visit to a fresh title hitches, then runs smoothly | Guides #161 and #290 on shader-compilation stutter | Here a first-time hitch is the topic itself; this guide takes only the cache-rebuilding test |
| How to pick a driver release in general, such as Game Ready against Studio, or an older release for an older card | #132, driver-release choice | This piece uses one move — putting the immediately previous release back in place a single time, as a test — and does not teach release choice |
| Laptop slow on battery, normal when plugged in | #362, laptop battery mode | This piece is written for a desktop pattern; a laptop needs its plugged-in against on-battery test run separately |
| One complete driver cleaning | #256, driver cleaning | This piece uses the rule that one complete cleaning is enough, and does not rewrite that guide |
Two rows in that table merit closer attention, since they are the rows readers most often confuse with this shape. The first is the Windows Update replacement. The trigger there is automatic: the system fits a driver on its own, the vendor switches to Microsoft, and the release steps back. The trigger here is manual: the owner fetched the newest vendor driver and fitted it by hand. The two triggers produce a similar complaint — titles run slower after a driver changed — but their checks run in opposite directions, and mixing them up points the reader at the wrong set of checks. The second is the wake-triggered fault. A machine that is slow only after waking from sleep, and normal again after restarting the title, has a wake trouble, not an update trouble. The update in that shape is a coincidence, and the roll-back test would prove nothing, because the fault follows waking rather than the release.
When to Stop Testing
Halt once the seven checks have named no layer. A title that has been pinned to one stretch, weighed against the release notes, replayed through the same stretch a trio of times, run through a roll-back test, release digits weighed on both sides, run through a settings restoration, item by item, and run through a range test — and still yielding no clear answer — has exhausted what home testing can offer. Pushing on from there is guesswork, and guessing at driver versions that were never at fault only clouds the picture. Restore the options changed for the tests before halting, so the machine is left in a settled state.
Halt when the roll-back, the settings put back, and the cache rebuilt are all finished and the title is still slow, and the temperature or fan figures are off, or picture corruption, a black screen, or a crash has appeared. Those signs sit with another family of guides, and further gaming tests now risk turning a testing question into a hardware fault. Write out the fresh sign, when it appeared, and which test was running at the time, and carry that note along. That note forms part of the evidence, and it tells more than any list of options that were changed.
Halt also when a driver test has been run a single time and the outcome stayed where it was, and a second option has been changed and the outcome stayed where it was. Two tried factors and an outcome that has not shifted is an answer by itself: those factors do not govern the layer. Trying a third, fourth, and fifth driver release one after another gains nothing. It leaves a pile of changes to unwind before an inspection can begin at all.
Keep the baseline notes once testing halts. The clock, wattage, and temperature figures, the card usage, the frame-time graph, the 1% low figure, and the frame-rate figure, and the driver release digits on both sides of each test, written out in the same stretch before and after each test, together carry more weight than any single one of them. An inspection begins with those notes, which is why the first check asks for the baseline to be written out before any test runs.
What to Bring for Inspection
Carry the notes from the seven checks: the pinned stretch and the baseline figures; the driver release digits ahead of the update and behind it, and ahead of the roll-back test and behind it; whether the release notes listed the same symptom; the three frame-rate figures from the first-outing test; the outcome of the roll-back test; which option, if any, brought the frame rate back when it was restored one at a time; whether only one title or every title is slow; and the clock, power, and temperature figures, including whether the clock rose as the load rose.
Carry the title name, the stretch that was pinned for the tests, and the options that were in use, because the control-panel energy preference and the in-game quality options are part of the evidence, and the outcome of restoring them is part of the diagnosis. Carry the card model and driver release, with the processor model, since the pairing forms part of the evidence. Leave the options as they are now, with every test option restored, and do not carry a list of each option changed for a test and then restored. The evidence is the current options plus the seven test outcomes. Past changes count only where one of them shifted the frame rate, since that change names the layer.
For inspection, HK GPU Fix is at Dali Plaza 1209, Shek Mun. The phone number is 2186 8010 and WhatsApp is 6558 6806. The published inspection fee range is HK$400 to HK$2,800, and no price is estimated before inspection. There is no Mong Kok shop and no on-site service.
Frequently Asked Questions
I installed the newest driver by hand and my game is slower. Is the new driver definitely at fault?
Not necessarily. Putting in a driver alters more than one thing at once: it swaps the driver files, and it can also rewrite control-panel choices, reset profiles kept per game, alter the power preference, reset the display refresh rate, and wipe the stored shader cache. Any one of those alterations can slow a title down. The rollback test is the check that pulls them apart: put the previous release back in place, retest the same spot for the same sixty seconds, and see whether the frame rate comes back. If it does, the release layer is the answer. If it does not, the release itself is not the culprit, and the diagnosis moves to the choices, the cache, and the power plan.
My game stutters only the first time I enter a new area after the update. Is that a regression?
Most often it is the shader cache being rebuilt. Putting in a new driver wipes the old cache, so the first entry into a fresh spot must compile shaders again, and frame times can leap while that happens. Play the same spot three times and note the frame rate each time: if it gets better each time, the cache was being rebuilt, and any judgement must wait until the cache is built. If the slowness is the same on all three runs, the cache is not the explanation. An opening-run stutter must never be written up as a lasting loss of speed.
I rolled back to the old driver and the game is still slow. Does that mean the old driver is broken too?
No. It means the release was not the culprit, and something the update altered besides the release is still in place. The installer can reset the control-panel power preference, overwrite profiles kept per game, alter the Windows power plan, or reset the display refresh rate — and a rollback puts the release back without putting those choices back. Look at each choice one at a time, put back one item, retest, and note which item, if any, brings the frame rate back. That item names the layer.
How can I tell a version fault apart from a settings fault?
By testing them as two separate variables. The rollback test tests the release alone: only the release alters, and everything else stays untouched. The choices-restoration test tests the choices alone: one choice is put back at a time, with a fresh test after each change. If the frame rate comes back on the old release, the release layer is confirmed. If the title stays slow on the old release and comes back only after one choice is put back, the choices layer is confirmed. Altering both at once and then deciding that the rollback failed proves nothing, because the test cannot tell which alteration mattered.
My graphics card clock is stuck low after the update. Does that mean the card is faulty?
Not by itself. The clock tracks demand: when the card is idle, a low clock is an unremarkable response, and the boost figure printed on a specification sheet carries no pledge that the card sustains that speed at every moment. Read a low clock alongside the usage, wattage, and temperature figures, and follow it as the load shifts. A clock that rises as the load rises describes a card sitting idle. A clock that refuses to budge however heavy the load grows describes a card being pinned down, and only that shape implicates the power-state layer. One clock figure cannot diagnose anything.
Only one game is slower after the update. Is that a driver problem?
It can be, but the shape points at the title or engine layer first: the new driver may not suit some function of that one title, or the title itself may need an update. Test other titles the same way. If only one title is slow and the rest run normally, the fault sits in the title or engine layer, and the next move is to look for a title update and the title's own choices. If every title is slow, the fault sits in the driver or system layer, and the diagnosis goes back to the release, choices, cache, and power-plan layers.
Should I stay on the old driver version permanently?
No. A rollback used in this piece is a test, not advice to stay on the old release for good. Staying on the older release while watching the release notes for a hotfix is a reasonable holding position. A newer release may gain a hotfix later, and the release notes are the place to watch for one — but no date can be stated, because no vendor has promised one. Do not write the old release up as always better than the new one: a regression can affect one title, one API, or one pairing, while the new release runs other titles well.
How many times should I clean and reinstall the driver?
A single time, fully. One full cleaning and one clean reinstall is enough. Repeating the same cleaning process does not change the outcome; it only hides the true cause by adding more changes to undo. If the roll-back test, the settings restoration, and the cache rebuilt are all finished and the title is still slow, the next move is not another cleaning. It is to halt the testing, restore the options changed for the tests, and carry the records for inspection.
I rolled back, but the driver version changed again on its own. What happened?
Windows Update can put a newer driver release back in place on its own after a rollback, and that quietly undoes the test. The release number must be noted down before the rollback test and noted down again afterwards. If the two numbers differ, the test is void, because the driver under test is no longer the one intended. Run the test again with the numbers checked before and after. This is a guard on the test, not the topic of this piece: the mechanism of a driver replaced by Windows Update is covered in its own guide.
When is it time to stop home testing and book an inspection?
Halt when the roll-back, the settings put back, and the cache rebuilt are all finished and the title is still slow, and the temperature or fan figures are off, or picture corruption, a black screen, or a crash has appeared. Carry the notes from the seven checks: the pinned stretch and the baseline figures, the driver release digits ahead of the update and behind it and ahead of the roll-back and behind it, whether the release notes listed the same symptom, the three frame-rate figures from the first-outing test, the roll-back outcome, which option if any brought the frame rate back, whether one title or every title is slow, and the clock, power, and temperature figures — together with the title name, the card model, the driver release, and the processor model, and those notes, plus the options as they stand now, are where an inspection starts.
Sources
The sources set out below were drawn on for this guide. Repair guides are drawn on for how the fault works and how it is tested, not as recommendations, and any figures from them are quoted as source claims where the text says so.
- tier1settings.com, "How to Roll Back a GPU Driver (NVIDIA, AMD, Intel)" — https://tier1settings.com/how-to-roll-back-gpu-driver/ — a new driver can regress, with lower average frame rate, worse 1% lows, shader-cache stutter, flickering, or crashes; a rollback is reversible; the Device Manager roll-back option only keeps the immediately previous version, and a clean installation or a driver-cleaning tool removes the old package and leaves the roll-back option greyed out; the release notes Known Issues or Open Issues should be checked; hotfixes are stated by that site to usually arrive within one to three weeks, which is that site's claim and not a vendor promise; a change under around 3% is stated by that site to fall within run-to-run variation, which is that site's claim and not a general standard.
- umatechnology.org, "After Graphic Driver Update Game Runs Badly" — https://umatechnology.org/after-graphic-driver-update-game-runs-badly/ — a checklist for a game running badly after a driver update; a rollback is presented by that site as the fastest repair, while this piece treats it as a test; where a rollback does not help, a clean installation is the next move; a driver-cleaning tool is presented as a thorough cleaning for advanced use.
- driver-updater.org, "GPU stutters after Driver Update" — https://driver-updater.org/en/wiki/problems/grafikkarte-ruckelt-nach-treiber-update/ — a micro-stutter must be pulled apart from a whole frame-rate fall before diagnosing; frame-time tools include Afterburner with RTSS, CapFrameX, and PresentMon; Windows Update can put an older driver release back in place on its own.
- rottenwifi.com, "Shader Cache Causing Stutter or Crashes?" — http://rottenwifi.com/shader-cache-is-making-a-game-stutter-or-crash-fixes-to-try/ — a new driver can invalidate or wipe the old cache, the opening runs must compile shaders again, and the stutter improves afterwards; a damaged cache can make a title crash or run worse, but not every stutter is a cache fault.
- windowsforum.com, "NVIDIA Shader Cache Size" — https://windowsforum.com/news/nvidia-shader-cache-size-reduce-windows-11-game-stutter-and-frame-time-spikes.448609/ — the NVIDIA shader cache has a limit, and exceeding it evicts older shaders so older games must compile again; the control-panel cache-size options include Disabled, 128 MB, 256 MB, 512 MB, 1 GB, 5 GB, 10 GB, 100 GB, and Unlimited.
- umatechnology.org, "Fix Game Stutter by Increasing Shader Cache Size" — http://umatechnology.org/my-games-stuttered-until-i-changed-one-shader-cache-setting-nobody-talks-about/ — NVIDIA states that putting in a new driver wipes the shader cache, the opening runs stutter, and that is an expected cache rebuild; the AMD shader-cache default is described as AMD optimized.
- umatechnology.org, "Windows Update Replaced My Graphics Driver" — https://umatechnology.org/windows-update-replaced-my-graphics-driver/ — Windows Update can swap a vendor driver for a Microsoft release and make performance worse; a rollback and hiding the update is that guide's topic, and it is used here only to guard the rollback test, not as this piece's topic.
A title that ran normally before a hand driver update and runs slower after it is not a verdict on the new driver. Fix one spot, read the release notes, play the same spot three times, run the rollback test with the release number checked on either side, and put the choices back one at a time. Those tests tell whether the release, the choices, the cache, or the power plan is holding the title back — and they tell it before any hardware is altered.