Prime95 can use AVX, AVX2 and, on supported CPUs, AVX-512 instructions. These wide instructions do more work per clock, and they usually draw more power and produce more heat than ordinary code. Knowing what they do, and how to switch them off, lets you choose between a brutal worst-case test and a gentler, more realistic one.
- AVX, AVX2 and AVX-512 are CPU instruction sets that process many numbers per instruction. Prime95 uses the fastest one your CPU supports.
- They raise power draw and heat, and many CPUs lower their clocks under heavy AVX load.
- You can disable them with flags in
local.txt. The file only exists after you run Prime95 once and close it. - Turning AVX off isn’t always cooler. On a CPU already pinned at its power limit, the result can be the same.
What are AVX, AVX2 and AVX-512 in Prime95?
They’re sets of CPU instructions that work on several numbers at once. Instead of multiplying one pair of numbers per instruction, an AVX instruction can handle a whole row of them, and each newer generation widens the row. Prime95’s maths is a perfect fit, because its fast Fourier transforms are huge batches of identical arithmetic. If you’re unsure why that arithmetic matters, see Small FFTs vs Large FFTs vs Blend.
| Instruction set | Idea | Typical effect in Prime95 |
|---|---|---|
| SSE2 | Older, narrower vector instructions | Slower, cooler baseline |
| AVX | Wider vectors than SSE | Faster, more power |
| AVX2 | Wider integer and fused-multiply support | Faster again, higher power |
| AVX-512 | Widest vectors, on supported CPUs only | Highest throughput and power draw |
According to forum discussions of Prime95 versions, AVX optimisations arrived after version 26.6, and AVX-512 support arrived in version 29.8, with the release notes covering the change history. Current builds, such as 30.19 build 20, automatically choose the fastest instruction set your CPU supports.
Why AVX makes Prime95 hotter
Because the CPU does more work in every clock cycle, it draws more power, and power turns into heat. Many processors also respond to heavy AVX loads by lowering their clock speed, so the total effect is “more work per cycle, fewer cycles per second, more watts overall”. Forum users have long noted that Prime95 versions after 26.6 run hotter because of AVX, and that the extra heat is much higher than most real programs produce.
If your cooler or power delivery can only just handle everyday loads, AVX-heavy Prime95 may overwhelm it. That’s a legitimate finding, but you should know whether you’re testing “my PC under normal use” or “my PC under the worst load imaginable”.
What real tests show
The effect isn’t uniform, and the best published example shows why. When SkatterBencher tested a Core i7-13700K at stock settings, turning AVX2 off made no difference to the average Prime95 result. Both runs averaged about 253.0 W and 90°C, because the chip was already pinned at its 253 W power limit. The limit, not the instruction set, set the outcome (SkatterBencher #50).
Their overclocked configurations behaved differently. For example, in their “Unleashed Turbo” setup, AVX2 on averaged 288.9 W at 100°C, while AVX off averaged 268.2 W at 95°C. That’s a meaningful difference once the chip has room to draw more power.
| Core i7-13700K, Prime95 Small FFTs | AVX2 on | AVX off |
|---|---|---|
| Stock (253 W limit) | 253.0 W, 90°C | 253.0 W, 90°C |
| Overclock: Unleashed Turbo + XMP | 288.9 W, 100°C | 268.2 W, 95°C |
Source: SkatterBencher #50, on an EK custom water loop. One CPU and one cooler, so don’t generalise the exact figures.
The lesson is that AVX matters most when your CPU has headroom to spend. If a power limit or a thermal limit is already the bottleneck, turning AVX off may change little.
How to disable AVX or AVX-512 in Prime95
The mechanism is a settings file called local.txt in the Prime95 folder. There’s one catch: Prime95 only creates it after you’ve started the program at least once and closed it completely. Then you add lines for the instruction sets you want to turn off.
- Run Prime95 once and close it fully. Make sure it isn’t still running in the system tray.
- Open
local.txtin a text editor. Back it up first. - Add the flags you want. A value of 0 disables that code path, and 1 allows it. Guides such as Igor’s Lab list these:
CpuSupportsAVX512F=0 CpuSupportsAVX2=0 CpuSupportsAVX=0 - Save and restart Prime95, then start the torture test.
The flags aren’t fully documented in one place. Read undoc.txt in your Prime95 folder for the exact names that apply to your build before relying on them.
Which setting should you test with?
| Goal | Setting | Why |
|---|---|---|
| Find your cooling or power ceiling | Full AVX (default) | It’s the harshest load. If it passes, you have a lot of margin. |
| Check stability of everyday clocks | AVX2 only, or AVX off | More realistic. Avoids failures that only occur under extreme AVX loads. |
| Avoid the extra heat from AVX-512 | AVX-512 off | Useful if your CPU supports it and you don’t use it day to day. |
| Test the exact load your software uses | Match your workload | If your renders use AVX2, test with AVX2 on. |
If you test both ways, you learn whether a failure only shows up under the heaviest instructions. That’s useful information. It might mean your overclock is stable for normal use but not for AVX-heavy work, and an AVX offset in the BIOS (below) could fix it.
Pitfalls and Linux notes
- Edits not applied. The file must exist (run the program once) and Prime95 must be restarted.
- Typos. A wrong flag name is simply ignored, so verify against
undoc.txt. - Linux (mprime). An Intel community thread describes a segmentation fault on Linux with a Core i9-14900K when AVX2 was enabled. That report is anecdotal and doesn’t establish a general fix, but disabling AVX2 is a reasonable troubleshooting step if you see a crash.
- Don’t mix up errors. If disabling AVX removes errors, your system may be unstable only under AVX load. See Prime95 errors explained.
AVX offsets in the BIOS
Many motherboards offer an AVX offset or AVX ratio setting, which lowers the CPU’s multiplier while AVX instructions run. It’s a way to keep a fast everyday overclock stable and cool while still allowing heavy AVX loads to run slower. If Prime95 with AVX is unstable but non-AVX tests pass, an AVX offset is the usual remedy. SkatterBencher’s article mentions such offsets (for example an AVX negative ratio offset in one configuration), which shows how common they are on enthusiast boards. Check your motherboard manual for the exact name and range.
Always monitor temperatures while testing, using the approach in safe Prime95 temperatures.
A timeline of instruction-set support (from the release notes)
The Prime95 release notes give a dated trail of when each instruction set arrived. This is more reliable than forum memory.
| Version | What the notes say |
|---|---|
| 27.1 and 27.3 | Add AVX optimizations for 32-bit and 64-bit builds. |
| 28.3 | Adds FMA3 support, which the notes tie to the Haswell CPU. |
| 28.6 | Adds AVX2 trial factoring support, with the notes citing about a 50% speedup over the previous version. |
| 29.1 | Adds AVX-512 trial factoring support. |
| 29.8 | Adds AVX-512 FFT support. The torture test dialog options are now based on cache sizes, and torture tests that use all RAM are now more stressful. |
Notice what that says about AVX in the stress test itself. AVX arrived in the FFT code in the 27.x series, FMA3 followed in 28.3, and AVX-512 FFT support arrived in 29.8. If you read an old forum post saying Prime95 runs cool, it was almost certainly written before those changes. Note that trial factoring is a different part of Prime95’s work from the FFT-based torture test, so only the FFT entries bear directly on stress-test heat.
Why CoreCycler defaults to non-AVX Prime95
A telling data point on testing choices comes from the CoreCycler README. The author says they’ve defaulted CoreCycler to Prime95 without AVX and AVX2 and using “Huge” FFTs. The README also says testing with SSE is lighter and can reveal boost-related instabilities, while noting that AVX and AVX2 should also be tested for full coverage. In other words, a respected single-core testing tool treats “AVX off” as the better default for finding boost instability, and “AVX on” as an extra layer on top. That matches the idea in this guide: AVX on and off answer different questions.
A simple test matrix
| Run | AVX setting | What a pass tells you |
|---|---|---|
| 1 | AVX off (or AVX2 off) | Your clocks and voltage are stable under a wide, high-boost load. |
| 2 | Default (AVX on) | Your cooling, power limits and voltage hold up under the heaviest instruction mix. |
| 3 | Real workload | Your actual applications are stable. Nothing replaces this. |
If run 1 passes and run 2 fails, an AVX offset in the BIOS or a lower power limit is the usual fix. If run 1 fails, the problem is more basic: core voltage, clocks or memory. See Prime95 errors explained.
Frequently asked questions
Should I disable AVX when stress testing?
Disable it for a more realistic or cooler test. Keep it on for the worst-case power and heat check, but make sure your cooling can handle it.
Does Prime95 support AVX-512?
Yes, on CPUs that support it. AVX-512 support in Prime95’s FFTs arrived in version 29.8, and current builds use it automatically.
Where is the Prime95 local.txt file?
In the Prime95 folder. It’s created after you run the program once and close it completely.
Why does my CPU downclock in Prime95?
Heavy AVX loads draw a lot of power, so many CPUs reduce clock speed to stay within power and temperature limits.
Will turning off AVX always lower temperatures?
Not always. On a CPU already at its power limit, the result can be identical. It helps most when the CPU has headroom to draw more power.
Is disabling AVX cheating the test?
No. It simply changes what you’re testing. Disabling AVX tests a more typical load, while leaving it on tests the worst case. Both are useful for different questions.