Introduction
1.1 APPROACH AND LIMITS OF EVIDENCE
PDF · 3Page 3AT WHAT LEVEL ARE WE INVESTIGATING?
There are four levels at which to investigate memory behavior. Choose the method to suit the question: repeatable tests, technical documentation, BIOS code or signal measurements.
| 1. Hands-on testing | Change BIOS settings, repeat POST and tests. Check stability, performance, and temperature. |
|---|---|
| 2. Documentation and parameter readout | Compare specifications, other people's results, and parameters read from the system. Separate the meaning of a field from assumptions about how it works. |
| 3. BIOS/MRC and registers | Examine the code, registers, and logs of a specific firmware version. Establish which memory-training stage ran and what it changed. |
| 4. Physical measurements | Observe signals and timing intervals with an oscilloscope or analyzer. Interface measurements do not reveal every process inside the die. |
Readouts show the current system state. Establishing the cause of a difference requires repeated checks under controlled conditions. For signal measurements, account for timing, measurement error and the effect of the connection itself.
WHERE I WORK AND WHAT I CAN CLAIM
I work at the first two levels: testing system behavior and examining technical data. I assess individual settings by comparing repeatable results. When no difference is visible, I leave the question open.
MY EXAMPLE: READ TIMING CENTERING 2D
In a recent discussion, I examined Read Timing Centering 2D under Training Algorithms. It is a switch for a training stage, not a numerical timing such as tCL. The ASUS BIOS Manual [1][1] lists Auto/Disabled/Enabled without explaining the algorithm. The RDTC2D field in Intel FSP for Raptor Lake-S [2] likewise offers only enable/disable. This explains the field's purpose, but does not establish how a particular BIOS applies it. I prefer to leave it Enabled, but this is not a firm recommendation. We will examine each setting in detail in a new chapter in subsequent revisions.
1.2 WHO THIS GUIDE IS FOR AND HOW TO USE IT
PDF · 8Page 8You can read this manual from start to finish if you are just getting started, or use it as a reference if you already know the fundamentals. In both cases, the goal is the same: to understand why a setting is changed, how it relates to the other settings, and how to verify the result.
IF YOU ARE TUNING DDR5 FOR THE FIRST TIME
First, distinguish JEDEC from XMP and understand how capacity, rank, and topology affect frequency. Then establish a known-good baseline. Do not start with a full set of someone else's values: if something goes wrong, it will be difficult to find the cause. The links below provide a short tuning path.
IF YOU ALREADY HAVE EXPERIENCE
The manual can be used as a reference. Go directly to the relevant group of timings, voltages, or training parameters, but consult the fundamentals chapter whenever the platform, memory capacity, or module type changes. For LGA1851 and CUDIMM, use the relevant sections: the IMC, CKD and training differ from Z790.
HOW TO APPROACH TUNING
1. SAVE THE BASELINE
First test JEDEC/Auto, then test XMP separately. Record the BIOS version, actual frequency, voltages, and trained values.
2. CHANGE ONE THING AT A TIME
In each iteration, change one parameter or one predefined group of parameters. Even related voltages should initially be adjusted one at a time; change them together only once the role of each is understood.
3. REPEAT MEMORY TRAINING
After significant changes, check not only a warm restart but also a full cold boot. One successful boot does not yet prove repeatability.
4. USE MULTIPLE TESTS
No single test covers every type of error. Your test suite must check the modules themselves, the memory controller, and system behavior in real workloads.
5. LEAVE SOME MARGIN
The final daily profile should be slightly more conservative than the last borderline point and remain repeatable at a different temperature and after the system has been idle.
WHAT TO RECORD AFTER EACH STEP
For each journal entry, it is enough to record the changed parameter, frequency, main voltages, module temperature, test results, and cold-boot behavior. A short, clear entry is more useful than a large collection of unexplained screenshots: it lets you quickly identify the last successful point and determine which change introduced the error.
WHEN YOU CAN STOP
A profile can be considered ready when it is repeatable across several cold boots, passes a suite of different tests, and produces no errors in normal workloads at operating temperature. Running one test indefinitely provides less confidence than several workloads that exercise different parts of the system.