Stability-testing tools
7.1 What needs to be demonstrated
PDF · 65Page 65Use different workloads to check a DDR5 profile. Data storage and transfer, the memory controller, heavy computational workloads, training repeatability, and margin at real operating temperatures must be tested separately.
FIVE DIFFERENT TEST TARGETS
| Target | What may be unstable | How we test it | What a single test does not prove |
|---|---|---|---|
| DRAM and data | Cells, rows, refresh, writes, and reads | Karhu, TM5, MemTest86, OCCT Memory | IMC stability under AVX load |
| IMC / PHY | CPU-to-DIMM link, VDD2/MC, TX, SA, Vref, and signal integrity | y-cruncher VT3/FFTv4, OCCT CPU + RAM | Absence of a rare retention error |
| CPU and cache | Cores, ring/cache, voltage, and CPU temperature | y-cruncher SVT and small in-cache tests | Stability of the modules alone |
| Training | Variation in trained values, cold/warm boot, and the saved fast path | Several complete training cycles and boots | Stability of an already-booted system |
| Temperature | tREFI, tRFC, DRAM voltage, and different thermal behavior of 1R/2R | Karhu/TM5 with HWiNFO and real airflow | Behavior of a cold memory kit |
PROFILE TESTING SEQUENCE
| Step | Stage | Practical action | Proceed when |
|---|---|---|---|
| 1 | Quick screening | 10–30 minutes of a targeted workload after one change | No errors or hangs |
| 2 | Isolate the cause | Compare SVT with VT3/FFTv4 and an independent RAM test | The conditions that reproduce the failure are understood |
| 3 | Pre-boot | MemTest86: four passes after cold training | No address or bit errors |
| 4 | Thermal conditions | Bring the DIMMs to their daily-use temperature plateau | The profile holds its operating tREFI |
| 5 | Repeatability | Cold boot, warm restart, sleep, and normal workloads | The result does not depend on one lucky POST |
DOCUMENTATION REVIEWED: The methodology of this revision; Numberworld y-cruncher FAQ; Karhu RAM Test FAQ; PassMark MemTest86 User Guide; official OCCT test descriptions.
7.2 y-cruncher: component workloads
PDF · 66Page 66Use only Component Stress Tester. The official Windows build is v0.8.7.9547b; Linux is v0.8.7.9547. My screenshot shows the Windows binary 0.8.7.9547b.
WHAT EACH COMPONENT DOES
| Tag | Algorithm | Primary workload | Role in DDR5 diagnostics |
|---|---|---|---|
| BKT | Basecase + Karatsuba* | Scalar integer multiplication; small working set | CPU arithmetic; little dependence on RAM |
| BBP | BBP Digit Extraction | Vector floating-point arithmetic; small working set | CPU compute workload; does not test the full RAM allocation |
| SFTv4 | Small In-Cache FFTv4 | Small FFTv4: floating-point Fourier transform in cache | CPU/cache; complements large FFTv4 |
| SNT | Small In-Cache N63 | Small N63: integer NTT in cache | CPU/cache with a different type of arithmetic |
| SVT | Small In-Cache VT3 | Small VT3: vector computation with an in-cache working set | CPU/cache; I compare it with large VT3 |
| FFTv4 | Fast Fourier Transform v4 | Floating-point FFT; large working set, strongly dependent on RAM bandwidth | CPU/cache/IMC/RAM with heavy memory traffic |
| N63 | Classic 64-bit NTT v2 | 64-bit integer NTT; large working set | Another algorithm for combined CPU and RAM testing |
| VT3 | Vector-Scalable Transform v3 | Vector transform; large working set, CPU + IMC + RAM | Mixed workload across the DDR5 path; successor to VST |
* Command Lines.txt expands BKT to include Toom-Cook; the developer’s internals guide says Toom-3 was removed.
DOCUMENTATION REVIEWED: Alexander J. Yee: y-cruncher v0.8.7, Version History, Multiplication Internals, Memory Allocation Guide, and FAQ; official package 0.8.7.9547b.
7.3 KARHU RAM TEST
PDF · 67Page 67RAM Test is not obsolete merely because its interface is old. Claims like that, made without examining the algorithm, are the usual bait from people who do not understand how or why any of this works: by that logic, mathematics is obsolete too, since most of its algorithms appeared long before DDR5. To discuss specifics, show DDR5 instability scenarios that Karhu misses and comparative testing that proves it. RAM Test was developed and is published by the Finnish company Karhu Software Oy; the stable version is 1.1.0.0, while the 2.0 branch is in development.
HOW LONG TO TEST: CURRENT DEVELOPER STATISTICS
| Time to error | Errors found | How to use it |
|---|---|---|
| 10 minutes | 74.89% | Quickly exposes a badly chosen setting |
| 30 minutes | 87.62% | Intermediate check after a timing or voltage change |
| 1 hour | 92.97% | Minimum for finding some intermittent errors |
| 3 hours | 97.84% | Reasonable test for a finished candidate |
| 6 hours | 99.30% | Strong final run; returns diminish sharply after this |
| 12 / 24 hours | 99.91 / 99.99% | Occasional search for an intermittent failure, not a daily ritual |
SETTINGS AND THERMAL CONDITIONS
| Field | Practical value | Why |
|---|---|---|
| Megabytes | About 85% of physical RAM; leave roughly 15% for the system | Do not select all memory: Windows and background processes need some headroom, without paging |
| Threads | Number of physical cores; 24 for the 14900K/KS | As recommended by the Karhu README |
| CPU cache / RNG | Default / Default | Set CPU cache to Enabled only for a combined CPU/cache test; XORWOW is primarily needed for Windows 7 |
| Stop on error | Enable | The first error already invalidates stability |
DOCUMENTATION REVIEWED: Karhu Software Oy RAM Test 1.1.0.0: official page, README, current FAQ, and archived 2021 FAQ; HWiNFO monitoring. Screenshot: my test run.
7.4 TESTMEM5 / TM5
PDF · 68Page 68
The original TestMem5 0.12 was written by Serj; CoolCmd continued its development. The current version, 0.13.1, fixes premature completion, an incorrect tested-memory size, and silent allocation errors. Run it as administrator and verify the TM5.exe/TM5.dll version.
THE CONFIG DETERMINES THE NATURE OF THE TEST
| Config | Creator | Purpose | Role for Intel DDR5 |
|---|---|---|---|
| DDR5 Intel / OptimalforIntelDDR5 | anta777 | Profile specifically for Intel DDR5 | The most logical starting point |
| Universal 2 | LMHz | Modern general-purpose profile; default in 0.13.1 | Quick general check |
| Absolut | anta777 | Profile focused on testing the memory array | Final independent check |
| Extreme | anta777 | Older, long-running general-purpose profile | Usable, but not DDR5-specific |
| 1usmus v3 | 1usmus | Short diagnostic profile | Historical; do not treat it as the DDR5 standard |
| Default | Serj | Original baseline sequence | Not the best choice for modern overclocking |
KARHU AND TM5 APPROACH A SIMILAR TASK DIFFERENTLY
| Characteristic | Karhu RAM Test | TestMem5 0.13.1 |
|---|---|---|
| Model | Single closed algorithm | Engine with user-defined configs |
| Progress | Time and coverage | Specified duration, sequences, and tests |
| Configuration | Cache, RNG, FPU, memory size, and threads | Patterns, blocks, duration, and order |
| Repeatability | Easier to compare identical runs | The result depends more heavily on the config |
| License | Paid | Free |
DOCUMENTATION REVIEWED: CoolCmd/TestMem5 v0.13.1 release and documentation; bundled configs by Serj, LMHz, anta777 and 1usmus; 02.09.2026.
7.5 MEMTEST86 FREE / PRO
PDF · 69Page 69PassMark MemTest86 v11.7 Build 1000 is a standalone UEFI application. Write the image to a USB drive, select the drive in the Boot Menu, and MemTest86 then runs without Windows or Linux. It is used for basic and service diagnostics of memory kits in UEFI mode.
CONFIGURATION FOR DDR5
| Setting | Practical choice | Meaning |
|---|---|---|
| CPU Mode | Parallel | Fastest standard mode; Single/Round Robin/Sequential are only for diagnosing UEFI or CPU cores |
| Passes | 4 passes | Free: maximum of 4; the first is shortened, the next 3 are full. Four full passes are available only in Pro/Site through PASS1FULL=1 |
| Tests | Free: 0–10, 13; Pro/Site: 0–13 | Tests 11/12 are 64/128-bit tests available only in Pro/Site; Test 13 Row Hammer is available in Free |
| Test 14 DMA | Disabled by default | Additional DMA/USB diagnostics, not a standard memory-overclocking test |
FREE, PRO, AND ECC
| Feature | Free | Pro |
|---|---|---|
| Core tests | Up to 4 passes; up to 16 logical CPUs; without Tests 11/12 | Up to 999 passes; more CPUs; Tests 11/12 available |
| Automation | Manual launch and standard report | mt86.cfg, profiles, advanced reports, and cache policy |
| ECC | Detection and reporting of supported system ECC | ECC injection on compatible CPU/IMC combinations as well |
DOCUMENTATION REVIEWED: PassMark MemTest86 v11.7: User Guide, editions, execution time, ECC, Test 13/14. SK hynix DDR5 32Gb M-die Rev.1.2, §4.37. Photo: my v11 Pro run.
7.6 RAM TEST PRO
PDF · 70Page 70RAM Test Pro 1.5.0 is a free 64-bit memory-testing utility for Windows 10/11 by Svetlana (PCSTONKS). It combines configurable test sequences, error logging, WHEA monitoring, NUMA/Large Pages support, and a separate DDR2–DDR5 benchmark.
WHAT SETS RAM TEST PRO APART
| Component | What is implemented | Practical meaning |
|---|---|---|
| Engine | Its own x64 core and cfg format, according to Svetlana's description | Flexible configuration of test sequences and parameters |
| Algorithms | Inverse, mirror move, refresh, pseudorandom, and I/O/AVX | The sequence and parameters can be changed through a config |
| Diagnostics | Error location, log, WHEA, and stopping at a limit | More context than a single Errors counter |
| Platforms | Multicore CPUs, NUMA, and experimental Large Pages | Useful beyond a typical gaming system |
| Benchmark | Read/Write/Copy, latency, and random access | An additional comparison point, but not a replacement for AIDA64/MLC |
DOCUMENTATION REVIEWED: Svetlana / PCSTONKS: RAM Test Pro 1.5.0 official page, guide and release notes.
7.7 Additional tools
PDF · 71Page 71These programs are optional, but they help quickly confirm a specific hypothesis and provide a second opinion. Karhu + y-cruncher remain the foundation of my methodology; Svetlana’s (PCSTONKS) RAM Test Pro is covered in 7.6 as a free alternative for an additional check. Always save the version, settings, duration, temperature, and first error together with the screenshot.
OCCT MEMORY CONFIGURATION EXAMPLE
| Tool | When useful | Limitation | Practical duration |
|---|---|---|---|
| OCCT Memory | Quickly catches transfer, frequency, and timing errors under Windows | Cannot see memory occupied by the operating system | 15–30 minutes |
| OCCT CPU + RAM | Tests the overall CPU and memory path; monitors WHEA | More difficult to isolate the cause | 20–45 minutes |
| HCI MemTest 7.0 | Simple alternative Windows RAM test | More manual setup; slower than Karhu. Pro makes thread allocation easier | 400%+ as a basic reference point |
| Real-world workload | Game, archiver, compilation, rendering, and system sleep | Does not provide controlled coverage | Normal work cycle |
7.8 RECOMMENDED TEST COMBINATIONS AND DURATIONS
PDF · 71Page 71| Stage | Primary combination | What it covers | Guideline |
|---|---|---|---|
| One timing | Karhu | Quick feedback on the latest change | 10–30 minutes |
| Frequency / power | y-cruncher SVT + VT3, then FFTv4 | CPU/cache versus IMC/RAM and a large working set | 20–30 minutes |
| Candidate profile | y-cruncher + Karhu; TM5/RTP are only additional cross-checks | Different algorithms under Windows | 45–60 minutes + 1–3 hours of Karhu |
| Pre-boot | MemTest86 4 passes | Gross address, bit, and Hammer errors | Until completion |
| Final daily profile | MemTest86 + y-cruncher + Karhu + cold/warm boot | Data, IMC, temperature, and repeatability | 4 passes + 45–60 min + 3–6 hr of Karhu |
DOCUMENTATION REVIEWED: OCCT official test descriptions; HCI MemTest 7.0 official page; PCSTONKS RAM Test Pro documentation; consolidated methodology of this manual.
7.9 Diagnosing errors and deciding when a profile is ready
PDF · 72Page 72After an error, first establish the conditions under which it appears. Examine the test, time, temperature, WHEA, boot type, and latest change; then revert only one parameter.
SYMPTOM -> FIRST HYPOTHESIS -> NEXT TEST
| Symptom | First hypothesis | What to do next |
|---|---|---|
| The data-error counter rises immediately | Latest OC change; CPU/cache, memory, IMC, or signal integrity | Revert the step; repeat from a known-good baseline; save the error text and conditions |
| The test cannot allocate memory | Requested allocation, program limitations, or available RAM | Reduce the allocation; check program settings and messages |
| The test runs unusually slowly | Paging (page transfers between RAM and disk), background load, or test mode | Check RAM/page-file use and repeat under comparable conditions |
| Error only after warm-up | tREFI/tRFC, DRAM VDD, cooling, or 2R | Record the SPD Hub temperature; lower tREFI or improve daily-use airflow |
| SVT and VT3 fail | CPU/cache, ring, CPU power, or an overall CPU/IMC, thermal, or power limit | Back off the CPU OC; compare SFTv4/SVT and OCCT CPU |
| SVT passes, but VT3/FFTv4 fail | IMC, VDD2/MC, TX/SA, memory, or signal integrity | Compare with Karhu; restore Auto ODT; check the voltage ranges |
| MemTest86 reports errors before the operating system loads | Gross memory/IMC instability, addressing, or UEFI/BIOS | JEDEC, another DIMM/slot, then repeat after Clear CMOS and cold training |
| Error only after cold boot | Non-repeatable training result or accelerated boot path | Disable MRC Fast Boot; perform three cold boots and compare runtime readback |
| WHEA without a RAM error | CPU/IMC/interconnect, not necessarily DRAM | OCCT CPU + RAM and a separate CPU test; check the WHEA log |
Data errors and memory-allocation errors are counted separately. Paging (transferring memory pages between RAM and disk) can dramatically slow the test; it is not a sufficient explanation for a rising RAM-error counter.
COMPLETION CRITERIA FOR A DAILY PROFILE
| Step | Requirement | What to save |
|---|---|---|
| 1 | Four error-free MemTest86 passes | HTML/photo of the report and the 11.x version |
| 2 | 45–60 minutes of y-cruncher SVT + VT3 + FFTv4 | Version, memory allocation, limit, and initial/final temperature |
| 3 | Karhu for 3–6 hours; TM5/RTP only as an additional run | Duration, coverage, memory size, threads, and 0 errors |
| 4 | Three cold boots and three warm restarts; sleep if used | POST behavior, Q-Code, and trained values |
| 5 | Final thermal run with daily-use cooling | Maximum SPD Hub temperature, WHEA, and ambient temperature |
| 6 | Normal workloads show no regressions | Archiver/game/rendering and a backup of the profile |
DOCUMENTATION REVIEWED: Consolidated methodology of chapters 03-07; official y-cruncher, Karhu RAM Test, PassMark MemTest86, OCCT and HWiNFO documentation.