- Can the cooler keep the die at a stable temperature?
- Can the power delivery hold up under a sudden all-core load?
- Does the silicon still compute correct answers at the clocks and voltages the firmware is using?
Passing the first says almost nothing about the third. This guide walks through a staged routine for Windows 11 and Linux, with a pass condition for each stage. It also corrects a few details that circulate with this kind of advice.
Current tool versions
Check these against the vendor pages before you download anything.
- y-cruncher: The developer's site lists version 0.8.7 Build 9547, released November 23, 2025, for both Windows and Linux. A later "b" Windows-only build recompiled the launcher without static linking, because the static version triggered virus-scanner false positives. If you see a DLL error on Windows, you may need the MSVC redistributable. The y-cruncher tool's own changelog also notes that the stress tester now displays elapsed time when an error occurs. That is useful for logging.
- OCCT Personal: OCBASE's download page lists v17.1.10 as the current stable build. That page is not among the y-cruncher results, so treat 17.1.10 as the build seen at the time of writing and recheck it.
- Prime95: GIMPS lists version 30.19 build 20 on its download page. Use whatever build the page offers when you read this.
Why the order of tools matters
No single program exercises everything.
| Tool | What it's best at | Main caveat |
|---|---|---|
| Prime95 | Worst-case heat and power draw. It verifies its results and logs failures. | It is harsher than almost any real application, so a fail needs interpretation. |
| OCCT | Cooling and power behavior, with live graphs and error flags. | Many modes, so there is a learning curve. |
| y-cruncher | Memory-path and instruction-kernel errors. It checks its own results. | Text-driven and less visual. |
| stress-ng (Linux) | A scriptable load with a verify switch. | Needs separate tools for sensors and logs. |
Prime95's Small FFTs mostly stay in cache, which means maximum heat with little memory traffic. Blend mixes cache and RAM load. Large FFTs lean on the memory controller. Treat these as practical tendencies, not guarantees that one mode isolates one component.
On temperature limits, use your own CPU's number. AMD lists a 95°C maximum operating temperature and a 120 W default TDP for the Ryzen 7 9800X3D. That figure is specific to that chip. Look up the maximum for the processor you actually own.
Steps 1 to 3: Prepare the machine
Step 1: Record what you're testing. Note the CPU model, BIOS version, memory speed and cooler. On Windows, run this in an elevated PowerShell window:
Get-CimInstance Win32_Processor | Select-Object Name, NumberOfCores, NumberOfLogicalProcessors, MaxClockSpeed
Get-CimInstance Win32_BIOS | Select-Object SMBIOSBIOSVersion, ReleaseDate
Get-CimInstance Win32_PhysicalMemory | Select-Object Manufacturer, Capacity, ConfiguredClockSpeed, PartNumber
On Linux:
lscpu | grep -E "Model name|Core|Thread|MHz"
sudo dmidecode -s bios-version
sudo dmidecode -t memory | grep -E "Size|Speed|Part Number"
Photograph the BIOS screens that hold your tuning: memory profile, power limits, Curve Optimizer or voltage offsets, and fan curves.
Step 2: Return the BIOS to a known state. Load the default profile and test once with no memory profile. Then test again with your intended daily configuration. If the default run passes and the tuned run fails, your changes caused it. If both fail, suspect cooling, power or a hardware defect.
Note that board makers set power limits differently. Some boards let the chip draw well above the vendor's guidance, which turns a cooler test into a PSU test. Intel says the Core Ultra 200S Plus series launched on March 26, 2026. It says existing 800-series motherboards support the chips, with DDR5-7200 memory support. Don't assume every board enables those memory settings by default, and you don't need them for an ordinary CPU stress test.
Step 3: Set up logging. On Windows, run HWiNFO in sensors-only mode and enable CSV logging at a one-to-two-second interval. Watch package temperature, hottest core, package power, effective clock and any throttling rows. On Linux:
sudo apt install -y lm-sensors stress-ng
sudo sensors-detect --auto
sensors
Use dnf with lm_sensors on Fedora. You should see a k10temp (AMD) or coretemp (Intel) reading. If you don't, fix that before you continue. A stress test with no temperature feed is a gamble.
Steps 4 and 5: Baseline and smoke test
Step 4: Idle baseline. Leave the machine alone for ten minutes. Record idle temperature, package power and fan speed. This gives you a reference for how far the CPU climbs under load. It also shows whether the machine settles back afterward.
Step 5: Five-minute smoke test. Run the OCCT CPU test, or Prime95 with Small FFTs, for about five minutes. Watch the temperature, not the clock.
- Pass: temperature climbs in the first minute, then plateaus below your limit with no errors.
- Fail: temperature keeps climbing, or jumps to the maximum within seconds. That usually means poor cooler contact.
Stop immediately if temperatures get within a few degrees of the CPU's rated maximum. Solve the cooling problem before any longer test.
Step 6: Prime95 torture test
- Launch Prime95 and choose Just Stress Testing.
- Set the thread count to your logical processor count.
- Choose Small FFTs for a heat and power run, or Blend for a stability verdict.
- Start HWiNFO logging, then start the test.
- Run at least 30 minutes for the main pass, and watch the first three minutes closely.
Reading failures:
- One worker fails repeatedly while others pass: one core is marginal. That often points to an undervolt or Curve Optimizer offset that is too aggressive on that core.
- Every worker fails within seconds: suspect memory settings or a power problem.
- A worker stalls but the window stays open: check for thermal or power protection in your sensor log.
A Prime95 failure is meaningful evidence that the configuration isn't stable under that workload. It doesn't prove the chip is defective. Small FFTs run hotter than nearly any real software. If Small FFTs fails but Blend, OCCT and y-cruncher pass for hours, you are making a judgment call about margin.
A laptop caveat that needs precision
Dell's support article describes a Prime95 problem on specific laptops. It covers the Dell Pro Precision 7 14 PW714260 and 7 16 PW716260 running Windows 11. Running Prime95 Blend or large FFTs can trigger a VIDEO_MEMORY_MANAGEMENT_INTERNAL stop error (0x0000010E). Dell's resolution is to use the small or smallest FFT settings for stress testing on those machines. Dell says the operating system's memory management is working as designed, and says NVIDIA recommends small FFTs too.
Two corrections to claims that circulate about this article. The Dell page we reviewed shows a last-modified date of 24 May 2026, not October 30, 2026. Our research notes list March 31, 2026 for another version, so confirm the date on Dell's page if it matters to you. And the problem is not a general "laptops can't take Prime95" rule, and it doesn't point to a faulty CPU. On any laptop, keep runs short, keep the charger connected, and check the vendor's notes first.
Step 7: Cooling and power with OCCT
Install OCCT Personal (free for personal use, not for business use) and open the CPU test. Set a 30-minute duration and let the instruction set stay on automatic unless you have a reason to change it. Pick a data set size: small stays in cache for maximum heat, larger pulls in memory traffic. Labels vary between builds, so read the tooltips in your version.
Run a steady load first. If your tune relies on an undervolt, repeat with a variable or cycling pattern. These catch fast current swings that a steady load can miss.
Watch these indicators:
- Temperature: it should rise and flatten. A curve still climbing after ten minutes means cooling never reaches equilibrium.
- Clocks: a staircase down as temperature rises suggests thermal or power-limit throttling.
- Power: package power should sit near the limit your BIOS applies.
- Error counter: any nonzero value is a failure, even if the test keeps running.
OCBASE describes a red exclamation mark as a critical compute error or a test crash. Reports differ by edition. Personal reports are image-based, and CSV and other formats depend on paid editions. Don't plan on CSV output from the free tier. Treat any single curve shape as a clue, not a diagnosis, because firmware limits and sensor reporting also affect readings.
OCCT also has a power test that loads CPU and GPU together. Run it only after the CPU test is clean, because a PSU near its limit can reboot the machine without warning.
Step 8: y-cruncher for quiet errors
y-cruncher checks its own results, so a wrong answer fails verification even if the PC never crashes. MajorGeeks describes the practical use this way: if a system passes lighter stress tests but fails y-cruncher, you probably have a RAM or memory controller issue. That is a rule of thumb, not a certainty.
- Download the archive, extract it and run the executable.
- Choose the stress-test entry, then the component stress tester. Menu numbers differ between releases, so read the menu text.
- Select the kernels, memory use and duration.
- Leave a few gigabytes of RAM free for the OS and monitoring tools. Otherwise Windows pages to disk and you are testing your SSD.
- Run the default set for 30 minutes. If you use an XMP or EXPO profile, run again with it enabled.
If it fails, note the kernel name and pass. A failure only in memory-heavy kernels points at memory speed, timings or memory-controller voltage. A failure in an instruction-set-specific kernel, with clean memory tests, points at core voltage or per-core offsets.
Step 9: stress-ng on Linux
stress-ng --cpu 0 --cpu-method matrixprod --verify --tz --timeout 30m --metrics-brief
watch -n 2 sensors
A clean run ends with a "successful run completed" message and exit status zero. Check echo $? straight after if you script it. Run at least two CPU methods, such as matrixprod and fft, since a core can tolerate one instruction mix and fail on another. Confirm the exact flags against your installed version's man page.
For a memory-focused variant, cap the share of RAM, for example --vm 2 --vm-bytes 75%. Stop if the system starts swapping or the kernel log shows out-of-memory kills.
Steps 10 to 12: Logs, verdict and sign-off
Step 10: Check the OS for hardware errors. A tool can report zero errors while the OS recorded faults. On Windows:
Get-WinEvent -FilterHashtable @{LogName='System'; ProviderName='Microsoft-Windows-WHEA-Logger'; StartTime=(Get-Date).AddHours(-3)} -ErrorAction SilentlyContinue | Format-List TimeCreated, Id, LevelDisplayName, Message
On Linux:
journalctl -k --since "3 hours ago" | grep -Ei "mce|machine check|hardware error|thermal|throttl"
No output is the result you want. Any WHEA event during a run counts as a failed configuration, even a corrected one. Corrected errors mean the hardware caught and fixed faults, which suggests you are at the edge. The event text often names the failing core.
Step 11: Translate the logs.
| What you see | Likely cause | First thing to try |
|---|---|---|
| Temperature never flattens | Cooler contact, pump or airflow | Remount cooler, check headers |
| Clocks drop as temperature rises | Thermal throttling | Improve cooling or lower sustained power limit |
| Clocks sag at moderate temperature | Power limit or VRM/PSU | Check BIOS limits and VRM airflow |
| One worker fails | Marginal core | Reduce undervolt on that core |
| Every worker fails fast | Memory profile or voltage too aggressive | Return to defaults, add tuning back one setting at a time |
| Only memory-heavy tests fail | Memory speed or timings | Lower speed or loosen timings |
| Hard reboot, no event | PSU or board protection | Test with a known-good PSU, lower power limits |
Step 12: Change one thing and retest. Restart only the failing test until it passes, then rerun the full plan once. Keep a log of the date, setting, test, duration and result.
Sign off when:
- Every stage completed with zero errors.
- The system logs show no hardware error entries.
- Temperatures plateaued with margin below the rated maximum.
- The machine returned to its idle baseline.
Repeat the short version (smoke test plus one Blend run) after any BIOS update, cooler change, memory change or large driver update. A clean result is strong evidence, not proof. It shows only that this machine, with these settings, completed these tests for these durations without observed errors.
About the logger script
The original guide includes a Python logger that wraps a load command and issues a PASS/FAIL verdict. I would not use it as an authoritative test as written. Three problems stand out:
- In monitor-only mode, with no load command, no temperature readings and no useful clock samples, it can still print PASS. Missing evidence should yield an inconclusive verdict.
- Its default 90°C limit doesn't match its own 88°C examples, and no single cutoff suits every CPU. Set the threshold from your chip's rated maximum.
- On Windows, the HWiNFO CSV is read only once after the run, so it doesn't monitor temperature during the test. psutil's clock readings can also be nominal rather than live.
Use HWiNFO or OCCT graphs on Windows and read the results yourself. On Linux, stress-ng's exit code plus the kernel-log check above gives a more trustworthy verdict.
Common mistakes
- Using one tool for ten minutes and calling it done.
- Changing the memory profile, fan curve and voltage in one go.
- Ignoring corrected WHEA entries.
- Trusting a single temperature sensor. A cool package reading can hide a hot VRM.
- Letting a memory test consume all RAM, so Windows pages to disk.
- Testing on battery, or with the side panel off in a cold room, then expecting the same result in a closed case in summer.
- Skipping backups. A crash can interrupt a disk write.
Quick answers
- How long? Five minutes catches obvious cooling faults. Thirty to sixty minutes per tool catches most stability problems. Add a multi-hour soak if you tuned voltages or memory. These durations are guidance, not a certification.
- Can it damage the CPU? A chip within its rated limits is operating as designed, and modern CPUs throttle or shut down to protect themselves. The real risks are excessive manual voltage, a failing cooler mid-test, or ignoring temperatures near the maximum.
- XMP or EXPO on or off? Test both. If the run passes without the profile and fails with it, the memory tuning is the culprit. The profile-enabled run is the one that matters for daily use.
The routine costs about an hour. Once you have logs from a known-good state, any later BIOS update, cooler swap or new memory kit gets a quick before-and-after comparison.
References
- CPU Stress Test: 12 Steps, 60 Min (2026) - shattered.io shattered.io · 2026-10-31T09:24:09+00:00
- Dell Pro Precision Laptops Using Prime95 Stop Responding with a 0x10E On-Screen STOP Error | Dell Australia dell.com
- Download y-cruncher 0.8.7.9547 - MajorGeeks majorgeeks.com