Does Prime95 Test Your GPU? The Common Mix-Up Explained

Prime95 is a freeware program by George Woltman for GIMPS, the Great Internet Mersenne Prime Search. Its Torture Test works by running FFT calculations on the processor and comparing every result against a known correct answer. Smallest and Small FFTs keep their data in the CPU caches, Large FFTs push out into system memory and stress the memory controller and RAM, and Blend mixes the two. Every one of those paths is processor and memory work. The graphics card sits idle through all of it, apart from drawing the screen. A system can pass a full day of Prime95 and still have a faulty GPU, and a system can fail Prime95 in minutes with a perfect GPU installed. Keeping those two statements separate will save you from the most common wrong turn in PC troubleshooting: replacing or clearing a component that the test never touched. Related platform notes are collected on the platforms hub.

Why people expect a GPU test in the first place

The confusion has understandable roots. Stress testing is often discussed as one general activity, and suites exist that bundle CPU, memory and graphics tests behind a single menu, so a famous stress testing name gets assumed to cover everything. Prime95 is also surrounded by the wider GIMPS world, where graphics processors genuinely do heavy mathematical work, and the two ideas blur together in forum posts. Add the fact that a Prime95 run heats the whole case, raises fan speeds everywhere and can expose a marginal power supply that a game then trips over, and it is easy to walk away believing the graphics card was involved in the test itself. It was not. Heat spreading through a shared chassis and load on a shared power supply are side effects, not coverage. The test’s own documentation draws the boundary cleanly, and this page simply follows it.

GPU computing in GIMPS is a different set of programs

Here is the part that makes the mix-up half true, and it is worth stating precisely. GIMPS, the prime search Prime95 was written for, can use graphics processors for real work: PRP testing and factoring stages can be carried out by separate GPU programs, of which gpuowl is the best known example, written specifically to run that mathematics on a graphics card. Those programs are not Prime95 and not mprime, they are not included in the Prime95 packages, and running them is participation in the prime search, not a graphics stability test with a pass-or-fail verdict about your card’s health. The most visible proof that GPU prime searching is real came in October 2024, when the largest known prime, 2^136,279,841 minus 1, was found by Luke Durant using cloud NVIDIA GPUs, as reported in GIMPS’s official news that month. None of that changes what the program on this site does on your desk: Prime95 tests CPU and RAM, and GPU prime-search software is a neighbouring project with a different purpose, a different download and a different definition of success.

What kind of tool actually tests a GPU

If the component in question is the graphics card, reach for a tool of a different type. GPU stress tools work by driving the card with continuous rendering or compute loads: sustained 3D rendering workloads that keep the graphics processor, its own memory and its cooler at full effort while watching for visual artefacts, driver resets, crashes and runaway temperatures. Monitoring software plays the supporting role, reporting GPU temperature, clock behaviour and power so you can see whether the card throttles or fails as heat builds. The selection principle is the same one this site applies to CPUs: match the tool to the component, run it long enough for heat to soak in, and record what happens instead of relying on a two-minute impression. What this page will not do is crown a specific product or quote version claims for GPU tools, because they sit outside the verified scope of this site; the type of tool and the signs to watch for are the reliable guidance. If your symptoms appear only in games, test the card under a sustained rendering load; if they appear everywhere, including on the desktop, work through the CPU and memory side as well rather than assuming one culprit.

How this fits a sensible troubleshooting order

Treat Prime95 and a GPU test as two instruments in one kit, used in the order the symptoms suggest. Random crashes under all kinds of load, errors in ordinary desktop work, or a worker stopping during a Torture Test point first at CPU, memory and their settings, and our guide to using Prime95 as a memory test explains how Large FFTs and Blend fit that investigation, including where a dedicated bootable memory tester takes over. Artefacts on screen, driver timeouts and failures that happen only in games or GPU rendering point first at the graphics card and its temperatures. Mixed symptoms justify both tests, run separately and recorded separately, so each result can be believed on its own terms. Short answers to questions of this kind, including what Prime95 tests, are also collected on our FAQ page.

Bottom line

Prime95 answers one question superbly: do your CPU and RAM keep calculating correctly under sustained load? It answers no question at all about your graphics card. GPU prime searching exists, but it lives in separate GIMPS programs built for that work, and GPU stability testing exists, but it lives in rendering and compute tools built for that component. Keep the three ideas apart, and every test you run will tell you something true. The wider platform picture is on the platforms hub.

Prime95 30.19 Build 20: What Changed in This Release

Prime95 is a freeware program by George Woltman for GIMPS, the Great Internet Mersenne Prime Search. A note on sources before the list itself: this page is drawn only from the 30.19 section of the official whatsnew.txt shipped inside the verified 30.19 build 20 packages. Items are paraphrased faithfully rather than decorated, and anything not in that section is not presented here as a 30.19 change. That discipline matters because version summaries on the wider web often mix changes from several releases into one list. One related detail worth knowing: the source package for this release is labelled build 21, while the distributed binaries are build 20. That difference is in the official distribution itself and is not a sign of a wrong download. Platform and version guidance across the site is indexed on the platforms hub.

The seven changes in 30.19

Read as a group, the pattern is clear. Four of the seven items improve the machinery of GIMPS contribution work: faster ECM curves, a more flexible ECM stage 2 that can continue from a GMP-ECM resume file, better runtime estimates for factoring stages, and easier manual setup of PRP work. Two items smooth communication and setup behaviour, proof uploads through Manual Communication and the worker preferences appearing when PrimeNet is enabled. One item, the timestamp option in undoc.txt, is a small power-user addition. None of the seven touches the Torture Test, the FFT presets, the AVX code paths or the stress.txt guidance, and this page will not pretend otherwise to make the release sound larger than it is.

How 30.19 relates to 30.18

The release immediately before this one, 30.18, recorded a single change in whatsnew.txt: a faster generic mod operation, noted as mainly benefiting external programs of the PFGW type that build on similar arithmetic. In other words, 30.18 was a one-line release and 30.19 is a seven-item release, and both are typical of this project’s mature phase, in which changes arrive as precise, technical adjustments to the prime-searching machinery rather than as interface redesigns. For a stress tester, the practical difference between the two builds is negligible; for a contributor, 30.19 is the more meaningful step, particularly for ECM work. The longer record, release by release and drawn from the same official file, is maintained on the version history page.

How to upgrade to 30.19 build 20

Upgrading follows the program’s portable design. First, download the verified package for your system from the download page and check its SHA-256 against the listed value, so the new build starts from a file you can account for. Second, protect your existing setup before changing anything: copy your current Prime95 folder, or at least its prime.txt settings file and your results files, to a safe place, because those files carry your worker configuration, any prime.txt options you have set, and the record of tests already run. Third, extract the new package, either into a fresh folder, moving your saved prime.txt and results files across, or over the existing folder if you prefer to keep one working directory; either way, confirm afterwards that your settings file survived with your options intact and that the program reports the new version when it starts. Fourth, run a short Torture Test before any long session, not because the release demands it, but because verifying a changed setup with a brief checked run is the same good habit this site recommends after any settings change. If anything looks wrong, your untouched backup folder restores the previous build exactly as it was.

Should stress testers bother upgrading?

Yes, as a matter of routine, with expectations set correctly. You gain a program whose change list you can read in full on one page, whose GIMPS machinery is current, and whose behaviour matches the documentation and guides written for 30.19, including every guide on this site. You should not expect a different stress testing experience: Small FFTs remains the maximum power and heat preset, Blend remains the default, preset sizes are still chosen from cache size as they have been since version 29.8, and the stress.txt recommendation of 6 to 24 hours for a thorough check is unchanged. Upgrading keeps your tools and your references in step with each other, which is reason enough, and it costs one verified download and a few minutes of care with two small files.

Bottom line

Version 30.19 build 20 is seven honest, technical changes, mostly in service of the prime search that is the program’s real job. Take the upgrade from the download page, keep prime.txt and your results safe while you do it, and expect your next Torture Test to feel exactly like the last one, which is precisely what a maintenance release should feel like. The full release record lives on the version history page.

Prime95 for Old Pre-AVX CPUs: Which Version Should You Use?

Prime95 is a freeware program by George Woltman for GIMPS, the Great Internet Mersenne Prime Search. Its Torture Test compares continuous calculations against known correct results, a method that works just as well on a fifteen-year-old processor as on a new one, because correctness does not age. What does age is everything around the method: instruction sets, operating expectations, and the features GIMPS itself now relies on. Owners of older hardware sit between those two facts, wanting a test that respects a pre-AVX processor without giving up a decade and a half of program development. Platform guidance for every generation is collected on the platforms hub.

Why version 26.6 is the one people search for

The timeline in the program’s own change history explains the attachment. Version 26.6 dates from around 2011. In version 27.1, FFT optimisation for AVX arrived, with further AVX work following in later releases, FMA support in the 28 series and AVX-512 FFT support in version 29.8. That makes 26.6 the last pre-AVX release, a matter of record rather than folklore. On top of that fact, a community tradition grew up around it: long-running enthusiast guidance, including the Tom’s Hardware Intel Temperature Guide as a community source, recommended 26.6 Small FFTs as a CPU heat test for older Core processors, arguing that its steady load sits closer to typical thermal design power and that later AVX versions could run considerably hotter on those platforms. That recommendation is community guidance about older hardware, not an official GIMPS position. The program’s author has never, in the included documentation, told users to downgrade to 26.6 for testing, and the included stress.txt guide describes the current Torture Test without any such exception. Both halves of that story deserve to be stated, because searches for 26.6 usually come from the community tradition, while the right modern answer comes from the program itself.

What 26.6 cannot do today

An old build is old in ways that matter beyond nostalgia. Version 26.6 predates the program’s modern PRP work and its proof generation, predates AVX-512 and heterogeneous processor support, and belongs to an era before the current settings layout, in which configuration moved to prime.txt in version 30.11. As a tool for contributing to GIMPS today it is simply out of date. As a pure heat generator it still runs, but heat generation was never the whole of what a Torture Test offers: the value of the modern test is that its presets are sized automatically from cache size since version 29.8, its results feed the current results format, and its behaviour matches the documentation and guides written for the current release, including everything on this site. Running 26.6 means testing with a program whose menus, defaults and documentation no longer match any current reference you will find when something goes wrong. Our dedicated page on Prime95 26.6 covers that release in full: why it is searched for, what it lacks, and how it compares with the current build. Note that this site does not offer a download link for the 26.6 binary; the page is an explanation and a guide back to the supported version, not a mirror of an obsolete package.

How the current version behaves on a pre-AVX processor

The current build does not require AVX to function. Its code includes paths for the instruction sets it can detect, and on a processor without AVX it calculates using the paths available to it. The Torture Test presets keep their official meanings on old hardware: Smallest FFTs test the L1 and L2 caches with high power and heat; Small FFTs test the caches with maximum power, heat and CPU stress for that machine; Large FFTs stress the memory controller and RAM; Blend, the default, mixes all of it. Two practical cautions apply to very old systems. First, keep expectations proportional: memory capacity, memory speed and cooling on a pre-AVX machine are all modest by current standards, so choose a writable folder, watch temperatures with whatever monitoring the system supports, and plan run lengths sensibly. Second, respect the official duration guidance in proportion: stress.txt recommends 6 to 24 hours for a thorough check and notes that unstable machines often fail within minutes, which is as true on a Core 2 system as on a new one. A short, watched first run remains the right way to meet an old machine before a long one.

Turning AVX paths off in prime.txt

On a processor from the early AVX era, or when comparing behaviour across machines, the modern equivalent of choosing 26.6 is to keep the current build and disable its AVX code paths deliberately. The documented settings, added to prime.txt in the Prime95 folder, are CpuSupportsAVX=0, CpuSupportsAVX2=0 and CpuSupportsAVX512F=0, each telling the program to act as if that feature is absent. On a truly pre-AVX processor these lines change little, because there is no AVX path to skip, but writing the intention into the configuration makes a test setup explicit and portable between machines. On an early AVX processor they reproduce the lighter, non-AVX load that made 26.6 popular, without abandoning the current program. The full method, including the move from the old local.txt name to prime.txt and what an AVX-off pass does and does not prove, is in our guide to disabling AVX in Prime95.

Bottom line

Old hardware deserves a clear answer rather than a museum piece. Version 26.6 earned its reputation honestly, as the last pre-AVX release and a community heat-test standard for its era, but the current build tests a pre-AVX processor correctly, documents itself properly and keeps you inside the supported program. Use 30.19 build 20, use prime.txt switches when you want an explicitly non-AVX run, and let the checked results, not the age of the binary, decide whether the machine is stable. The rest of the platform family is on the platforms hub.

Can You Run Prime95 on a Laptop? Power and Cooling Limits Explained

Prime95 is a freeware program by George Woltman for GIMPS, the Great Internet Mersenne Prime Search. Only the Windows version has a graphical interface; on a Mac laptop the same testing is done with the command-line mprime, and the physical limits described here apply in the same way. Nothing about a laptop makes the test invalid. The calculations are still compared against known correct results, so an error still means the machine computed wrongly under load. The honest question is not whether a laptop may run the test, but what a run on a thin, power-limited machine can and cannot prove. The rest of the platform family is indexed on the platforms hub.

The power wall comes first

A desktop processor under a Torture Test can often draw sustained power that its large cooler was chosen to handle. A laptop processor lives under a power wall set by the machine’s design: the adapter, the battery system, the voltage regulation and the thin chassis all share one tight budget. Under Small FFTs, which the program describes as maximum power, heat and CPU stress across the L1, L2 and L3 caches, a laptop will commonly reach that power limit quickly and settle at the clock speed the budget allows. That settling is normal engineering behaviour, not a fault and not a failed test. It does mean, however, that the run is validating the laptop at its sustained speed, which may be well below the brief boost speeds printed in marketing material. Judge the run by correctness and by whether temperatures settle, not by whether the clock stays at its highest number.

Cooling headroom is small, and heat soak is fast

The second limit is thermal mass. A laptop cooler has small fans, short heat paths and very little metal to absorb a burst of heat, and the keyboard deck and desk surface become part of the thermal system whether you plan for it or not. In practice, temperatures on a laptop climb faster and level off higher than on a comparable desktop, and a machine that is fine for the first ten minutes can be in a different state after an hour, once the whole chassis is warm. Dust in the vents, a soft surface blocking the intakes, or a hot room can each move the result. This is why the first laptop run should be short and watched: you are learning where this particular machine’s ceiling sits before you trust it with a long session. If temperatures climb and never settle, stop the test and deal with the physical side first. Our troubleshooting guide to a processor reaching its limit, Prime95 overheating and 100C readings, works through the checks that apply once a machine is at its thermal ceiling.

How to run a sensible first laptop test

Plug the laptop into mains power for the whole test; running a maximum stress load on battery drains it quickly and adds a second power limit on top of the first. Place the machine on a hard, flat surface with its vents unobstructed, and open a monitoring tool before you start so you can watch temperature, clock speed and throttling behaviour from the first minute. Choose Small FFTs for a first session of around ten to fifteen minutes, the short duration the official stress.txt guide reserves for exploring a step before committing to a long test. Watch whether temperature settles at a steady level and whether any worker reports an error. If the run stays correct and temperatures settle, you have a baseline. Longer validation, including Blend, the default preset that mixes cache and memory work, and Large FFTs for the memory controller and RAM, can then be planned in supervised sessions, working toward the official guidance of 6 to 24 hours only on a machine whose cooling has already shown it can settle. On many laptops, several shorter supervised runs with full records are more realistic than one unattended overnight run, and a well-recorded shorter run is better evidence than a long run nobody watched.

External cooling and everyday setup notes

Raising the rear of the laptop slightly or using a basic stand improves the airflow path under most machines, and an external cooling pad with fans aimed at the intake area can help a chassis that draws air from underneath. Treat these as airflow aids, not as transformations: they can lower the settled temperature and reduce throttling, but they do not turn a thin laptop into a desktop with a large tower cooler. Keep expectations in that proportion, and re-run the same short test after any setup change so the comparison is honest. Room temperature matters too; a test that settles acceptably on a cool morning may behave differently in a warm room in summer, which is worth noting in your record if the machine will be used for heavy work year-round.

What a laptop result does and does not prove

A correct, settled Prime95 run on a laptop proves the machine calculates correctly at the sustained speed its power and cooling allow, which is precisely the speed it will use for long renders, compiles and other heavy work. It does not prove the machine can hold its short boost clocks indefinitely, because no laptop is designed to do that under this class of load. An error, on the other hand, is meaningful on any machine: the result did not match, and the cause should be investigated at stock settings with clean vents and sensible ambient conditions before any conclusion is drawn about the hardware. For the broader question of what the program can and cannot do to a machine, and how monitoring fits into that judgement, read Is Prime95 Safe? alongside this page.

Bottom line

Run Prime95 on a laptop with respect for its two walls, power and cooling. Stay plugged in, keep the vents clear, start short with Small FFTs while watching the sensors, and extend the run only when temperatures have shown they can settle. Handled that way, a laptop test is genuinely useful: it validates real sustained behaviour instead of chasing a boost number the machine was never built to hold. More platform guidance is collected on the platforms hub.

Prime95 on Windows 10: Download and Run the Stress Test

Prime95 is a freeware program by George Woltman for GIMPS, the Great Internet Mersenne Prime Search. The Windows version is the only graphical version; Linux and macOS use a command-line program called mprime. Because this page exists to serve a specific search intent, it is worth saying plainly at the start what is and is not different on Windows 10: the download files are the same files, the presets are the same presets, and the way results are checked is the same method. Anyone trying to sell or promote a special Windows 10 build of Prime95 30.19 build 20 is describing something that does not exist in the official distribution. Platform guidance for every system is indexed on the platforms hub.

Which package does your Windows 10 system need?

If you are not sure which Windows 10 edition is installed, check the system information panel in Windows Settings before downloading; choosing by guesswork wastes a download and, on a 32-bit system, the 64-bit executable simply will not run. Most machines still in service run the 64-bit edition and should use the win64 package. The 32-bit package is a compatibility option for genuinely old installations, not a lighter version for modern hardware, and it should not be chosen on a 64-bit system for speed or caution reasons.

Download and verify

Get the package from the download page, which lists both Windows packages with their exact sizes and checksums, or use the official direct links: go/prime95-win64-official for the 64-bit file and go/prime95-win32-official for the 32-bit file, both pointing to the verified packages on mersenne.org. After downloading, compare the file’s SHA-256 with the matching value in the table above. A matching hash confirms the file is complete and unchanged from the verified official package. A mismatch means the download should be deleted and fetched again from the official link, not opened to see what happens. If a download arrives as an installer or under a different filename, treat that as a mismatch of a more obvious kind and start again from the download page.

Extract, launch and choose Just Stress Testing

Extract the full contents of the zip into a new writable folder, such as a Prime95 folder in Documents or on a second drive. Prime95 is portable: it stores settings in prime.txt and logs such as results.txt beside the executable, so the folder must remain writable and must not be a temporary extraction location or a protected system folder. Double-click prime95.exe when extraction is complete. Running as administrator is not required for a normal test, and there is no installation step, service or reboot involved. On first launch, choose Just Stress Testing if your aim is hardware validation rather than joining the GIMPS prime search through PrimeNet. The program will remember that choice in its settings file.

Run the Torture Test on Windows 10

Open Options, then Torture Test. The presets carry their official meanings: Smallest FFTs test the L1 and L2 caches with high power, heat and CPU stress; Small FFTs test the L1, L2 and L3 caches with maximum power, heat and CPU stress; Large FFTs stress the memory controller and RAM; Blend, the default, tests all of the above. Since version 29.8 the program chooses FFT sizes automatically from your CPU cache size, so ignore old tables that assign fixed size ranges to each preset. Watch temperatures with a monitoring tool while the test runs, and check your processor’s own specification for its temperature limit, because Prime95’s documentation publishes no universal number. For how long to run, the included stress.txt guide is the authority: it recommends 6 to 24 hours for a thorough check, observes that unstable machines often fail within minutes, and suggests short 10 to 15 minute runs only while stepping through overclock settings. Stop the test from the Test menu so the logs are preserved, and keep results.txt together with a note of the preset and duration.

Same test, same meaning as Windows 11

Results on Windows 10 mean exactly what they mean on Windows 11, which is the practical value of the shared package. A worker error on Windows 10 is a calculation that failed to match its known correct result under the tested conditions, and it should be handled the same way: note the details, return to conservative settings, and retest. A long error-free run is evidence of stability under that test on that machine, regardless of which of the two operating systems hosted it. If you maintain machines on both systems, you can use one written procedure, one set of presets and one results format for all of them. The companion page, Prime95 on Windows 11, walks through the identical flow written for that system’s search intent, and nothing in it contradicts this page, because there is nothing material to contradict.

Next steps

New testers should follow the full walkthrough in How to Use Prime95 for stopping a test and reading its outcome. Fresh verified packages are always on the download page, and the rest of the platform family, from Ryzen and Intel guidance to laptop and older CPU notes, is collected on the platforms hub.

Prime95 on Windows 11: Download and Run the Stress Test

Prime95 is a freeware program by George Woltman for GIMPS, the Great Internet Mersenne Prime Search, and the name covers the Windows graphical version of the program. Other systems use a command-line version called mprime, but Windows 11 gets the full graphical program described on this page. Its Torture Test stresses CPU and RAM by comparing continuous calculations against known correct results, which is why builders use it to validate a new PC, an overclock or a repair. This page covers the complete Windows 11 flow in order: get the exact file, verify it, extract it correctly, launch it the right way, and run a first test you can interpret. The platform overview lives on the platforms hub, and the Windows sections of the main download page are written for this same intent.

Step 1: Download the exact Windows 11 package

Use the official direct link go/prime95-win64-official, which points to the verified package on mersenne.org, or start from the download page, whose Windows 11 section lists the same file with its size and checksum. The file you should receive is p95v3019b20.win64.zip. If the filename differs, if the size is not 8,805,545 bytes, or if a site offers you an installer, a setup wizard or a Windows 11 edition of Prime95, stop and get the file from the link above instead, because the official program is distributed as this single zip package and has no installer. A 32-bit package also exists for old 32-bit Windows installations, but a normal Windows 11 system is 64-bit and should use the win64 package named here.

Step 2: Verify the file before you run it

Verifying the download takes a minute and proves the file arrived complete and unchanged. Open the folder containing the zip, then compare its SHA-256 with the value in the quick answer above: d9475f2ff3f4a6a701abc49a86a66126cb48abd10bda6fa87039d98fa8756bca. Windows can produce the hash from a command prompt or PowerShell window with its built-in checksum utility, and any reputable hash tool will do the same job. If the value matches exactly, the file is the verified official package whose hash was checked against the file on mersenne.org. If it does not match, delete the download and fetch it again from the official link rather than running a file you cannot account for. This one habit removes an entire category of doubt later, when a test result needs to be trusted.

Step 3: Extract to a writable folder

Right-click the zip and extract all of its contents into a new, writable folder, for example a Prime95 folder in Documents or on a data drive. Do not run prime95.exe from inside the zip preview, and do not accept an extraction wizard’s temporary folder as the program’s home. Prime95 keeps its settings in prime.txt and its logs, including results.txt, in the same folder as the executable, so the folder must stay writable and stay put. A folder inside a protected system directory is a poor choice because a normal user account may not be able to write there, which breaks the portable design. After extraction you should see prime95.exe together with its supporting library files and the included text documents, among them readme.txt and stress.txt. Keep the whole set together; moving the executable out on its own leaves required files behind.

Step 4: Launch and choose Just Stress Testing

Double-click prime95.exe. You do not need to right-click and run as administrator, and a normal stress test does not need an installer, a service or a reboot. On first launch the program offers a choice between joining GIMPS, where your computer searches for Mersenne primes and reports to the PrimeNet server, and just stress testing. For validating hardware, choose Just Stress Testing. Then open Options, then Torture Test, and pick a preset. Smallest FFTs test the L1 and L2 caches with high power and heat. Small FFTs test the L1, L2 and L3 caches with maximum power, heat and CPU stress, making that preset the right one for finding a temperature ceiling. Large FFTs stress the memory controller and RAM. Blend, the default, tests all of the above, and since version 29.8 the sizes are chosen automatically from your CPU cache. The full beginner workflow, including how to stop a test and read its outcome, is in How to Use Prime95.

Step 5: Monitor, run for a meaningful length, and record the result

Keep a monitoring tool open during the run so you can watch CPU temperature, clocks and power rather than guessing from fan noise. There is no single official safe temperature in Prime95’s documentation that fits every processor, so check the limit for your exact model on the manufacturer’s specification page. For duration, follow the one official source: the stress.txt guide included with the program recommends 6 to 24 hours for a thorough check, notes that unstable machines often fail within minutes, and reserves short 10 to 15 minute runs for exploring each step of an overclock before a long final test. When you stop, use Test, then Stop, so the logs are written cleanly, and keep results.txt with a note of the preset, the run length and your system settings. If a worker stopped on its own with an error, that is a real result: the calculation did not match the known correct answer under those conditions, and the next step is diagnosis at more conservative settings, not an immediate identical rerun.

Frequently asked questions

What is the exact Prime95 file to download for Windows 11?

The Windows 64-bit package for Prime95 30.19 build 20 is p95v3019b20.win64.zip, 8,805,545 bytes, with SHA-256 d9475f2ff3f4a6a701abc49a86a66126cb48abd10bda6fa87039d98fa8756bca.

Do I need to install Prime95 or run it as administrator on Windows 11?

No. Prime95 is portable on Windows 11: extract the zip to a writable folder and double-click prime95.exe. No installer and no administrator account are needed for a normal stress test.

Does Prime95 work differently on Windows 11 than on Windows 10?

No, there is no essential difference. The same Windows 64-bit package is used on Windows 11 and Windows 10, and the Torture Test presets, settings and result checking work the same way on both systems.

How long should I run Prime95 on Windows 11?

The stress.txt guide included with Prime95 recommends running the Torture Test for 6 to 24 hours for a thorough stability check. Unstable systems often fail within minutes, and short runs of 10 to 15 minutes are only for exploring overclock steps.

Next steps

For the complete workflow beyond Windows 11 setup, read How to Use Prime95. For the same flow on the older system, see the companion page for Windows 10, and return to the download page whenever you need a fresh, verified copy of the package. Platform-wide guidance, including AMD, Intel, laptop and older hardware notes, is indexed on the platforms hub.

Prime95 on Intel 13th and 14th Gen: How to Validate Stability

Prime95 is a freeware program by George Woltman for GIMPS, the Great Internet Mersenne Prime Search. Its Torture Test runs continuous calculations and compares each result with a known correct answer, stopping a worker when an answer does not match. That makes it a clear pass-or-fail instrument: at a given set of settings, either the system completes checked work without errors, or it does not. What a Prime95 result cannot do on its own is explain the internal reason a particular processor behaved a particular way, and on this platform that distinction matters more than usual. The wider platform context is collected on our platforms hub.

What Intel has said, in general terms

Intel has published statements addressing reports of instability affecting some 13th and 14th Generation desktop processors, and has released guidance that includes microcode updates delivered through system BIOS updates, together with a recommendation to use Intel Default Settings. That is the level at which this page will describe the situation, because it is what Intel’s public statements support. This page does not present a root-cause explanation, does not list specific voltage values, does not publish a list of affected models beyond the generation-level topic Intel itself addressed, and does not draw a conclusion about any individual processor, including yours. If you need the current official position, read Intel’s own statements and your motherboard maker’s BIOS notes directly, since those are the sources that can update their guidance over time, and a static guide page cannot speak for them.

Step one: update the BIOS before you test

Because Intel’s guidance is delivered substantially through BIOS and microcode updates, updating your motherboard BIOS to the latest release offered by your board maker is the first practical step, before any stress testing means anything. Follow the board maker’s own update instructions exactly, use a stable power source during the update, and note the BIOS version you installed. After updating, load Intel Default Settings in the BIOS rather than a saved performance profile, an automatic overclocking feature or a profile carried over from before the update. The point of this order is simple: a stability result only has value if you can state the conditions it was produced under, and current firmware plus the processor maker’s default settings are the cleanest baseline available. Testing performed on an old BIOS with an aggressive motherboard profile cannot separate firmware, settings and hardware questions from each other, so it wastes the one thing a careful test should produce, which is clarity.

Step two: validate with Prime95 at those defaults

With the baseline set, download the current version from the download page, extract it to a writable folder and choose Just Stress Testing on first launch. Open Options, then Torture Test. For a first validation, Blend, the default preset, is the sensible choice because it tests CPU, cache and memory behaviour together; the program describes Small FFTs as its maximum power, heat and CPU stress preset, and Large FFTs as the preset that stresses the memory controller and RAM. Since version 29.8, preset sizes are chosen automatically from cache size. Run with a monitoring tool open, watch that temperatures settle rather than climb without limit, and follow the duration guidance in the included stress.txt file: 6 to 24 hours for a thorough check, with the honest note that unstable systems often fail much sooner, sometimes within minutes. Record the preset, the BIOS version, the settings profile in use, how long the run lasted and the exact wording of any error from results.txt. A run you cannot describe afterwards is a run you cannot use, whether it passed or failed.

Reading the result without over-reading it

If Prime95 completes a long run without errors at Intel Default Settings on a current BIOS, you have solid evidence that the system, as configured, calculates correctly under sustained load. Keep the settings and the record. If a worker stops with an error at those same defaults, resist the urge to explain it yourself. An error tells you the system is not stable under the tested conditions; it does not, by itself, identify which component or mechanism is responsible, and the difference between those two statements is the whole discipline of this page. Our guide to Prime95 errors and what counts as unstable explains how to read error messages and separate a test result from a conclusion. Retest once after confirming the BIOS version and default settings were truly in place, because a profile that silently re-enabled a tuning feature is a common source of confusion. If the error repeats at verified defaults, stop tuning and move to the support step below.

The AVX offset, and why defaults come first

Intel platforms offer an AVX offset, a BIOS control that lowers clock speed when AVX instructions are detected, and it is a legitimate tuning tool on a system that is otherwise behaving as expected. Our dedicated guide to the Prime95 AVX offset explains how it works. On the systems covered by this page, however, an AVX offset should never be used to mask instability that appears at Intel Default Settings. Lowering AVX clocks to make an error disappear changes the conditions of the test; it does not resolve a system that fails at the defaults its maker recommends. Sequence matters: defaults and current firmware first, a clean validation second, and only then, on a system that passes, ordinary tuning decisions such as offsets, made one change at a time with a retest after each change.

If it fails at defaults: use Intel’s channels

A system that repeatedly produces Prime95 errors, crashes or other instability at Intel Default Settings on a current BIOS is a support case, not a tuning puzzle. Gather your records before you contact anyone: processor model as shown by your system, motherboard model, BIOS version, memory configuration, the Prime95 preset used, run duration and the exact error text. Then use Intel’s official support channels and, where applicable, its warranty process for boxed processors, or the system or motherboard maker’s support if your machine is a prebuilt system. Present the facts you recorded without adding a theory about the cause; support teams work fastest from clean, dated, repeatable information. Do not keep running long maximum-heat tests on a system you already know fails at defaults in the hope of producing a different answer. Once the pattern is documented and repeatable, additional stress adds heat and time, not information.

Bottom line

Update the BIOS, load Intel Default Settings, validate carefully with Prime95, and record everything. Pass at defaults and you have a trustworthy baseline for any later tuning. Fail repeatedly at defaults and the honest, effective route is Intel’s support and warranty channel, armed with the clear record your testing produced. That restraint is not a lack of ambition; it is how a stress test stays useful on a platform where official guidance, not community speculation, should lead.

Prime95 on Ryzen 7000 and 9000: Stress Testing Zen 4 and Zen 5

Prime95 is a freeware program by George Woltman, written for GIMPS, the Great Internet Mersenne Prime Search. Its Torture Test is a secondary use described in the included stress.txt guide: the program calculates continuously and compares each result against a known correct answer. That result checking is why it remains useful on a new platform: a Ryzen system can feel fast and still return a wrong answer under sustained load if voltage, memory or cooling margins are too thin. All platform guides on this site are collected on the platforms hub.

AVX-512 on Ryzen 7000 and 9000

Ryzen 7000 series processors, based on the Zen 4 architecture, and Ryzen 9000 series processors, based on Zen 5, support AVX-512 instructions. That matters for Prime95 because AVX-512 FFT support was added to the program in version 29.8, and the current 30.19 build 20 can therefore use a wide vector code path on these chips instead of falling back to narrower AVX2 work. In practical terms, a Torture Test run on these platforms is a genuinely heavy modern load, and heat and power behaviour during Small FFTs deserves respect from the first minute: a cooling setup that looks fine in short bursts may behave very differently once the whole package has soaked in heat for an hour. If you ever need to compare behaviour with the wide vector path removed, the program can disable instruction paths through prime.txt settings, covered in our AVX guidance linked below.

PBO, power limits and the temperature wall

Most Ryzen stress testing today happens under Precision Boost Overdrive, or PBO, rather than under a fixed manual clock. Under PBO the processor raises clocks while it stays inside its configured power, current and temperature limits, and it reduces clocks when it reaches one of those limits. That behaviour changes how you should read a Prime95 run. A falling clock speed during Small FFTs is not automatically a failure: it may simply be the processor managing itself against its temperature wall, the point at which it holds temperature by giving up frequency. What matters for stability is whether the calculations stay correct at whatever speed the processor settles into, and whether temperatures level off instead of climbing without settling. This is also why comparing two PBO configurations by clock speed alone misleads. A configuration that holds a slightly lower clock but completes hours of checked work is the better configuration, and a configuration that boosts higher for ten minutes and then produces a worker error has told you its margin is too thin. Record your PBO settings, your power limits if you changed them, and the ambient conditions alongside every serious run.

What Small FFTs means on Ryzen

The Torture Test presets are defined by where their data lives. Smallest FFTs test the L1 and L2 caches with high power, heat and CPU stress. Small FFTs test the L1, L2 and L3 caches and are described by the program as maximum power, heat and CPU stress. Large FFTs stress the memory controller and RAM. Blend, the default, tests all of the above. Since version 29.8 the FFT sizes for each preset are chosen automatically from your CPU cache size, so do not copy fixed size ranges from old forum tables. On a Ryzen build, use Small FFTs first to establish the honest worst case for core temperature and package power under PBO. Then use Large FFTs when memory speed or timings are part of what you changed, because Ryzen memory behaviour and core behaviour are tuned through related settings and a failure can live on either side. Finish with Blend for the long run, following the official stress.txt duration guidance of 6 to 24 hours for a thorough check, while remembering that unstable systems often fail within minutes and that short runs are only for exploring steps.

Monitoring the run properly

Prime95 itself is not a monitoring dashboard, so watch the run in a separate sensor tool. Many builders use HWiNFO, a free monitoring program from the HWiNFO website, to follow core temperatures, clock speeds, package power and throttling behaviour while the test runs. This site has no commercial relationship with HWiNFO; it is recommended because watching the sensors is the difference between a useful test and a guess. Watch for three patterns: temperature that settles, which is normal under a maximum load; temperature that keeps climbing, which points at cooling or airflow; and errors that appear only after heat soak, which point at a margin that disappears when the machine is hot. For a deeper reading, see our guide to Prime95 temperatures explained.

No AVX offset? The Ryzen way of thinking

Intel platforms offer an AVX offset, a BIOS setting that lowers the multiplier only when AVX work is detected. Ryzen platforms do not offer an equivalent Intel-style AVX offset, so the tuning levers are different. The same goal, keeping heavy vector loads inside a safe power and temperature envelope, is reached through the platform’s own controls instead: PBO power limits, temperature limits, curve and voltage behaviour, or, at the Prime95 level, disabling the AVX or AVX-512 code paths in prime.txt so a run can be compared with the full load. The principle is identical even though the menus differ: reduce the intensity of the heaviest load in a controlled way, then prove the result with a checked run, rather than accepting errors as the price of a headline clock speed. Our guide to the Prime95 AVX offset explains the Intel setting and the reasoning behind it, and that reasoning is the part that transfers to Ryzen even though the setting itself does not.

If you are shopping or studying alongside a build

Cooling is the component that decides most Ryzen stress test outcomes, and readers who are planning an upgrade around a high sustained load often look at large dual-tower air coolers such as the Noctua NH-D15, available via our Noctua NH-D15 link. If you are also deciding on a platform upgrade at the same time, start from the verified packages on the download page and the overview on the platforms hub before you change hardware and settings together, so you can still tell which change caused which result. For readers who want to study the underlying topics more formally, structured courses are available through our Coursera link. Disclosure: the two product and course links in this section are sponsored affiliate links. If you buy or enrol through them, this site may earn a commission at no extra cost to you. They are buying options, not requirements for running the test described above.

Bottom line

On Ryzen 7000 and 9000, let Prime95 do what it is built to do: run the current version so the AVX-512 path is in play, establish the temperature truth with Small FFTs, check memory with Large FFTs, prove the final PBO configuration with a long Blend run, and watch the sensors throughout. A clean, long, error-free run at settings you have written down is worth more than any short burst at settings you cannot reproduce.