How to Use Prime95: A Complete Guide
Prime95 runs heavy math workloads that simulate worst-case CPU stress. The program loads every core with operations based on the Lucas-Lehmer method, the same test used in the GIMPS project to check large Mersenne numbers. These calculations push power delivery, cooling, and memory pathways hard enough to reveal weaknesses in your hardware. Most users run Prime95 for stability checks, while others use it for benchmarks or to contribute to prime research.
The catch is that it pushes hardware harder than almost anything you would normally run, such as gaming or rendering, so starting without monitoring temperatures or knowing which mode to choose can lead to overheating or shutdowns. Here, we’ll explain how to run Prime95 safely, which Torture Test mode to choose, what to monitor while it runs, and when to stop.
What Prime95 Is Actually Doing
When Prime95 is running, it is actually searching for a rare type of prime number called a Mersenne prime, written as 2^p − 1 (where p is also a prime number). For example, when p is 5, the Mersenne number is 2^5 − 1 = 31, which is prime. Most Mersenne numbers are not, though. Out of millions tested, only 52 have ever been confirmed.
GIMPS runs each candidate through a series of stages, from quick checks that eliminate easy non-primes to slow tests that confirm the real ones. Prime95 handles most of these on your CPU, and each stage stresses your hardware differently.
| Stage | What Happens | Where It Runs |
|---|---|---|
| Trial Factoring (TF) | Quickly checks for small factors to filter out obvious non-primes | First filter, usually on GPUs |
| P-1 Factoring | A smarter method that catches larger factors. Uses a lot of RAM. | Second filter, runs on CPUs |
| PRP Test | A fast “probably prime” test using FFT-based squaring | The main test most users run today |
| Lucas-Lehmer (LL) | Confirms a positive PRP result is genuinely prime | Final verification step |
| ECM | Extra factoring for leftover composite numbers | Optional cleanup work |
The real stress comes from PRP and Lucas-Lehmer tests. Both rely on squaring giant numbers millions of times using a math shortcut called the Fast Fourier Transform (FFT). FFTs are extremely demanding calculations that push your CPU, cache, memory controller, and RAM hard all at once. Since the math allows zero mistakes, any unstable component, such as a weak overclock, low voltage, or overheating, gets caught quickly.
To understand why this puts such a heavy load on your CPU, look at the test at the heart of it all: the Lucas-Lehmer Test.
The Lucas-Lehmer Test
The Lucas-Lehmer test is Prime95’s core engine. It evaluates Mersenne numbers, a special class defined as 2^p−1. The test algorithm runs repeated squaring operations using modular arithmetic, which are extremely CPU-intensive and require precise calculations.
Because the workload stresses the processor, memory paths, and FFT performance, any instability, such as overheating, voltage drops, or degraded components, shows up quickly. This is why the Lucas-Lehmer test doubles as both a primality check and a reliable CPU stress method.
The math is simple to describe, but brutal for your CPU to actually run:
The numbers grow huge very quickly, so Prime95 uses Fast Fourier Transforms (FFTs) to keep the squaring process fast. In 2018, GIMPS added a newer PRP test that performs the same FFT work but includes built-in error checking, while the Lucas-Lehmer test is still used to confirm any positive result.
Modes of Operation
Prime95 takes the same heavy Lucas Lehmer calculations and turns them into tools you can use for real-world testing. It offers multiple modes designed for GIMPS work, stability checks, and hardware evaluation. Each mode runs a different mix of FFT sizes and memory usage, so you can target the CPU, the RAM, or both.
Torture Test
The Torture Test is the main tool for stability checks. It loads every CPU thread with FFT calculations that stress the core logic, cache, memory controller, and RAM. The mode you choose determines where the load is focused, so it helps to understand what each mode is designed for before starting a test.
The section below breaks down the available modes, what they stress, and when each one is the right choice:
Errors during Small FFTs usually point to CPU issues, such as low Vcore, overheating, or a weak core, while errors during Large FFTs or Blend typically indicate RAM or memory controller problems. A good approach is to run one to two hours of Small FFTs, followed by several hours of Blend. See our Prime95 troubleshooting guide for fixing specific errors.
Benchmarking
Prime95 also includes a Prime95 benchmark that measures how long your CPU takes to complete fixed FFT workloads. You get a clean, repeatable result to compare performance before and after tuning. Results are saved to results.txt, so you can track changes or improvements after overclocking, BIOS updates, cooler upgrades, or RAM adjustments.
Because the benchmark runs the same heavy math as the Torture Test, an unstable overclock will not just score lower; it will error out or crash, exposing weak tuning before it fails during real workload.
Performing CPU Stress / Torture Test
The Prime95 stress test pushes your processor and memory with a repeatable load that exposes instability quickly. Its Torture Test mode drives the CPU through intense integer-based FFT workloads while exercising the L1, L2, and L3 caches. Every core is used, which makes it useful after an overclock or undervolt.
This section shows how to run a torture test and monitor temperatures with Hardware Info. The goal is to help beginners run a proper stability check without confusion, using clear steps that cover FFT profiles, thread configuration, thermal safety, and real-time monitoring before and during the test.
Before You Start
Install Hardware Info or HWMonitor to watch temperatures, voltages, and clock speeds. Keep the sensor panel visible throughout the test. Stop immediately if temperatures rise past your CPU’s safe limit. Systems can fail from cooling issues, weak memory, or poor voltage control, so real-time monitoring matters. Things to check before starting:
- CPU temperature
- Any thermal throttling
- Sensor panel visibility
- Case and cooler performance
Never stress test without monitoring temperatures. Idle temperatures should sit around 30–45°C, while sustained load temperatures should stay near 85°C. Cross 95°C, and you risk thermal throttling at best and permanent CPU damage at worst.
Step-by-Step Guide
1. Launch and Choose Torture Test Mode
Open Prime95 and run the executable. When the window appears, select “Just Stress Testing” instead of joining the GIMPS project. You can also open the Torture Test screen from the Options menu. This keeps the program focused on system stability rather than prime-search tasks.
2. Select a test configuration
Prime95 opens two windows: one is the worker panel, and the other is the Torture Test selector. Each profile uses different FFT sizes; we use small FFTs as they create the heaviest CPU heat. Main options include:
Large FFT touches RAM, but it cannot replace a dedicated RAM test. These steps focus on CPU stability.
Prime95 auto-enables AVX instructions, which heat CPUs far beyond any real-world workload. Here’s how to disable it:
- New way (recommended): In the Torture Test dialog, uncheck AVX, AVX2, and AVX-512 before starting.
- Old way (if you don’t see those checkboxes): Fully close Prime95, go to the folder where you installed it, open local.txt in Notepad, add a new line that says CpuSupportsAVX=0, save the file, then reopen Prime95.
3. Configure Core Usage (Threads):
Prime95 will detect your threads automatically. The thread count usually equals twice your core count if Hyper-Threading or SMT is enabled. If you’re unsure, tools like HWInfo, HWMonitor, or CPU-Z show the exact number.
Verify that Prime95’s thread count matches your CPU to apply full-load stress across all cores.
4. Start the Test
Click OK to begin the test. Do not use your system for gaming, editing, or other heavy tasks during this period.
Prime95 loads your CPU aggressively, and some FFT phases produce more heat than others. Temperature spikes depend on cooler quality, airflow, room temperature, and which FFT size is currently running.
5. Monitor and stop
How long you run the test depends on what you’re testing:
Stop the test through the Test menu. Monitor temperatures in HWiNFO or HWMonitor while it’s running. Crashes, freezes, BSODs, or “FATAL ERROR” messages signal instability. A clean overnight run with no errors is a strong sign your CPU is stable for gaming, productivity work, and sustained heavy loads.
Conclusion
After using Prime95, what you do next matters. A successful Prime95 run does not guarantee the system will never crash, but a failed test is a sign that something in the setup is unstable.
However, if crashes, freezes, BSODs, or sudden restarts happen during tests, do not blindly increase voltage just to force stability, since that can create heat and long-term wear fast. Instead, fix the exact issue causing the failure, then test again properly. For specific errors and fixes, check our complete guide to fixing Prime95 not working.
Who Should Use Prime95?
Prime95 is free and works on Windows, macOS, and Linux, so anyone can run it. But now that you know how Prime95 works, the next question is if you should use it. Here’s who can:
- Overclockers: Verifying a new CPU clock speed or voltage is rock-solid before daily use.
- PC builders: Conducting stress-test a fresh build to catch hardware faults inside the return window.
- System troubleshooters: Confirming whether random crashes or BSODs come from an unstable CPU, RAM, or cooling.
- Cooling and case modders: Checking if a new cooler, fan setup, or thermal paste actually holds temperatures under full load.
- GIMPS contributors: Donating spare CPU cycles to the worldwide hunt for new Mersenne primes.
If you’re a casual user who browses, games, or works on documents, you probably don’t need Prime95 at all. Your CPU will never face this kind of load in normal use, and lighter tools like Cinebench or HWMonitor are usually enough.
