The three main tests in Prime95’s torture test window don’t measure the same thing. A CPU can pass hours of Small FFTs and still crash under Blend or Large FFTs, because each mode loads a different part of your system. This guide explains what an FFT is, why the size matters, what each mode actually stresses, and how to combine them so a pass genuinely means something.
- Small FFTs fit in CPU cache, so the cores do almost all the work. Best for core voltage, clock stability and cooling checks.
- Large FFTs spill into main memory. Best for RAM overclocks and the memory controller.
- Blend (the default) mixes sizes and uses most of your RAM. It’s the mode most likely to expose bad memory.
- Don’t rely on one mode. Run Small FFTs, then Blend, and read the result by which one fails.
What is an FFT in Prime95?
FFT stands for fast Fourier transform, a method for multiplying very large numbers quickly. Prime95 exists to hunt for Mersenne primes, enormous numbers of the form 2p − 1, and testing one means multiplying numbers with millions of digits over and over. FFT-based multiplication is what makes that practical.
That’s also why Prime95 makes such a good stress test. The maths has to be exactly right at every step, so if a core, a cache line or a memory cell misbehaves even once, the final answer is wrong and Prime95 notices. It’s a built-in error detector, running on the hardest workload it can find.
The FFT size is simply how big the data set is for each multiplication. You’ll see it in Prime95’s messages, written something like “1344K”. A larger FFT needs more memory to hold its working data, and that detail is the key to the three modes.
Why FFT size changes what you test
Your CPU has small, extremely fast memory built in, called cache, arranged in levels (L1, L2 and L3). Main memory, your RAM, is far larger but much slower to reach. Which one holds Prime95’s data depends on the FFT size.
- Small FFTs fit in the CPU’s cache. The cores stay busy and rarely wait for RAM.
- Large FFTs don’t fit, so data constantly moves between cache and RAM. The memory controller and RAM are now part of the test.
Cache sizes differ widely between CPUs, so the exact crossover point varies from one processor to another. The principle doesn’t change. Small stays in cache, large goes to RAM.
Small FFTs ask “can the cores do maths without a single mistake?” Large FFTs add “and can the memory system keep feeding them correctly?”
Small FFTs: the CPU heat and stability test
Small FFTs use data small enough to stay in cache, so main memory is accessed the least of any mode. With no waiting on RAM, the cores run flat out. That makes this mode the usual choice for finding an unstable core voltage or clock, and typically the one that produces the most heat and power draw on modern chips.
Use it for: checking CPU core stability after an overclock or undervolt, testing a new cooler, and finding your cooling ceiling.
Don’t rely on it for: RAM stability. Because it barely touches memory, it can pass while a memory problem sits undetected.
Newer Prime95 versions also offer a “Smallest FFTs” option, which uses an even smaller footprint and is the most extreme for heat and power. Confirm it exists in your version, and don’t leave it running unattended on a new overclock.
Large FFTs: the memory-heavy test
Large FFTs use data too big for the cache, so main memory is accessed heavily. PlayTool’s guide describes the “in-place” variant as one that uses FFTs too large for the cache while reusing the same small area of RAM, so it stresses the CPU-to-memory path under heat without testing a large amount of RAM.
Use it for: RAM overclocks, XMP or EXPO profiles, and the CPU’s memory controller.
Watch out for: attribution. An error here can come from the CPU, the memory controller or the RAM itself, so a failure doesn’t tell you which one to blame.
Blend: the all-rounder
Blend is the default for a reason. It cycles through a mix of FFT sizes from very small to very large, and it sets the memory to use to most of your installed RAM. So it tests the cores, the cache and a lot of the RAM in a single run. According to PlayTool, it’s the test most likely to fail if you have bad RAM.
The weakness is the mirror image of its strength. Because Blend covers so much, a failure doesn’t tell you which part is at fault. That’s why Blend works best after a clean Small FFTs run: if Small FFTs passed, a Blend failure points toward memory.
If your hard disk light stays on during Blend, it has been told to use more memory than your PC physically has and is swapping to disk. Lower the memory value in a custom test until the disk activity stops.
Side-by-side comparison
| Mode | Data lives in | Main target | Heat and power | Best for |
|---|---|---|---|---|
| Small FFTs | CPU cache | Cores and cache | Usually highest on modern CPUs | Core voltage, cooling |
| Large FFTs | Cache and RAM | Memory controller, RAM | Often lower than Small FFTs | RAM overclocks, XMP/EXPO |
| Blend | Cache and most RAM | Everything, mixed | Medium to high | General stability, bad RAM |
Summary based on the descriptions in PlayTool’s Prime95 guide and community discussions. Exact behaviour varies by CPU and Prime95 version.
Does Small FFTs really run the hottest?
On modern chips it usually does, but the old rule isn’t universal, and the evidence is a mix of ancient and awkward. The oldest data point comes from PlayTool, which measured a 3.2 GHz Pentium 4 with Hyper-Threading under Windows XP. On that system, In-place Large FFTs was the hottest single test, not Small FFTs.
| Test | Temperature |
|---|---|
| Idle | 34.3°C |
| Small FFTs | 48.8°C |
| In-place Large FFTs | 49.5°C |
| Blend | 49.3°C |
| MemTest86 (for comparison) | 47.3°C |
Source: PlayTool, “Torture test your CPU with Prime95”. One very old system, so treat it as a lesson about method, not a prediction for your CPU.
For a modern reference, SkatterBencher’s Core i7-13700K article logged Prime95 Small FFTs at stock settings at an average package power of 253.0 W and an average CPU temperature of 90°C, on an EK-Pro QDC Kit P360 custom water loop. Their overclocked runs reached about 289 W at 100°C (SkatterBencher #50). We didn’t run these tests ourselves, and a smaller cooler would produce higher temperatures.
Forum users give the same practical advice: Small FFTs is for peak heat and cooling, and Blend or Large FFTs is for stability across the whole system. One AnandTech poster explained the heat difference simply: Small FFTs run hotter because the cache is being used constantly, not because the maths is harder.
What AVX changes
Prime95 can use AVX, AVX2 and, on supported CPUs, AVX-512 instructions. These process many numbers per clock and usually draw more power, so they raise heat and often lower the clock speed the CPU can sustain. The effect is biggest in Small FFTs, where nothing else limits the cores.
Interestingly, SkatterBencher found that at stock settings, turning AVX2 off made no difference to the i7-13700K’s average result (253.0 W and 90°C either way). The chip was already pinned at its 253 W power limit, so the limit, not the instruction set, set the outcome. With overclocked settings, the AVX-off runs drew less power and ran slightly cooler. The takeaway is that AVX matters most when your CPU has headroom to spend. Our AVX settings guide shows how to switch instruction sets off.
Which mode fails first?
Forum reports, which are anecdotes rather than controlled tests, are fairly consistent. Small FFTs tends to throw errors or blue screens sooner when core voltage is the problem. Blend is slower to fail but better at catching memory trouble. Two examples: a user in an AnandTech thread reported passing Small FFTs and then hitting a 0x124 blue screen within two minutes of Blend, and a poster on Overclock.net described a system that was stable through hours of Small FFTs but crashed within half an hour of Large FFTs.
The lesson is the one in the lead: passing one mode says very little about the others.
How to choose and combine modes
| Your goal | Run this first | Then |
|---|---|---|
| Test a new CPU cooler | Small FFTs | Blend for a longer, steadier load |
| Check a CPU overclock or undervolt | Small FFTs | Blend, then a second tool |
| Check an XMP or EXPO memory profile | Blend or Large FFTs | MemTest86 outside Windows |
| General “is this PC stable” check | Blend | Small FFTs if Blend fails, to narrow it down |
The diagnostic logic is simple. If Small FFTs passes but Blend fails, suspect memory or the memory controller. If Small FFTs fails, look at core voltage, clocks and temperature first. If you only get one test, PlayTool recommends In-place Large FFTs because it runs hot while exercising the CPU and memory path. For the step-by-step process and run times, see how to stress test your CPU with Prime95.
Common mistakes
- Using Small FFTs as a RAM test. It barely touches memory.
- Calling a build stable after one mode. Run at least two.
- Ignoring temperatures. A pass at the thermal limit with heavy throttling isn’t the same as a clean pass. See safe Prime95 temperatures.
- Running Blend with too much memory. If the disk light is on, you’re testing your swap file.
- Treating one error as noise. An error means a wrong result under load. Here’s what to do.
Frequently asked questions
What is the difference between Small FFTs and Blend in Prime95?
Small FFTs fit in the CPU cache and stress the cores with little memory use. Blend mixes FFT sizes, including ones that use most of your RAM, so it tests the CPU, cache and memory together. Blend is better for finding bad RAM.
Which Prime95 test generates the most heat?
On modern CPUs, Small FFTs usually generate the most heat because the data stays in cache and the cores never wait on memory. On an old Pentium 4 test, In-place Large FFTs was slightly hotter.
Which Prime95 mode is best for testing RAM?
Use Blend or Large FFTs. Both push data through main memory. Blend is the test most likely to fail with bad RAM, but you may need another test to separate a RAM fault from a CPU memory controller fault.
Is Blend or Small FFTs better for overclock stability?
Neither alone is enough. Use Small FFTs for the CPU cores and Blend for memory and the whole system.
What does FFT stand for?
Fast Fourier transform, a method for multiplying very large numbers quickly. Prime95 uses it to test Mersenne numbers for primality.
Why does my CPU run cooler in Blend than in Small FFTs?
Blend spends part of its time waiting on memory, so the cores aren’t fully busy every cycle. Small FFTs keep data in cache and keep the cores continuously busy.