Reference · Overclocking
Glossary of the Bench
Every term on this desk is defined here once, in the sense this publication uses, with a note on where that sense differs from common use.
A machine that misbehaves answers in symptoms. A bench answers in words, and the two have to meet before any tool is opened. This page takes the eight terms this desk uses most, burn-in, stress test, stability, throttling, multiplier, base clock, offset and error correction, and states for each what it names, how it gets misread, and where the written record behind it stops.
| Term | What it names | Where the word gets misread |
|---|---|---|
| Burn-in | sustained operation over a stretch of time | read as a verdict when it is a longer sample |
| Stress test | a bounded run with defined patterns and a detection report | read as full coverage of a processor |
| Stability | behavior held inside the conditions a component is given | read as a permanent property of the part |
| Throttling | inconsistent performance | read as a slow part instead of a condition problem |
| Multiplier | bench usage: the factor applied to the base clock | read as freely available when vendors may lock it |
| Base clock | the reference frequency a multiplier works on | confused with the final clock it produces |
| Offset | a signed adjustment added to a reference value | read as linear when voltage heat rises quadratically |
| Error correction | machinery that catches memory errors before they spread | assumed present when the undetected errors are the silent kind |
The table is for consulting. The sections below are for the mechanisms, because a word used without its mechanism is how the misreads in the third column happen.
The words come before the tools
A diagnosis fails in the naming more often than in the measurement. A crash filed as a software bug gets a reinstall. The same crash filed as an unstable clock gets a longer run and a look at conditions. The record this desk draws on is blunt about the cost of the first filing: failures coming from an overclocked component might never be correctly diagnosed and can instead be attributed to bugs in applications, device drivers or the operating system. The right word buys something smaller and more ordinary. When the name is precise, the next step tends to choose itself, and the procedures behind it are already written on the pages covering load, burn-in and stability.
Burn-in and stress test: a duration and a question
A stress test is bounded. It loads a component with a defined set of patterns for a defined time and reports what it detected during that run, no more. A burn-in is a stretch of sustained operation, long enough that the slower kinds of failure have time to appear. The two overlap without coinciding, and the difference matters because of a coverage limit stated plainly in the record: software tools for testing hardware stability exist, and it is generally impossible for a private individual to thoroughly test the functionality of a processor. Every run is a sample. A pass is a statement about that sample, which is why the desk treats memory on its own page, where what a memory pass covers matters most. A burn-in does not escape the limit. It widens it, at the cost of the hours it consumes.
Why does a machine that passed one run fail on other software?
Because a pass is not a proof, and the mechanism has a name. Some manufacturing techniques, silicon on insulator among them, produce circuits with hysteresis: performance affected by past events. A particular sequence of state changes can work at accelerated rates in one situation and fail in another at the same voltage and the same temperature. A system built that way can pass a stress test and still be unstable in other programs. That is the ground stability stands on. It is not a property a component owns once. Operating voltage is often raised to maintain stability at accelerated speeds, and higher frequencies and voltages raise power consumption and heat with them, so every gain in stability is paid for in headroom. Bob Colwell, Pentium architect, compressed the arrangement into a phrase: "an uncontrolled experiment in better-than-worst-case system operation".
Throttling, and the other ways a machine says no
Throttling is the mild entry. It is inconsistent performance, speed that arrives and withdraws as conditions change. The rest of the list is harsher. An unstable machine can shut down or reboot without warning when heat, voltage or current pass what the system tolerates, and it can lose or corrupt data while appearing to work. Sustained overclocked use can also damage a component permanently enough that it misbehaves under normal conditions without ever becoming unusable, the hardest case to file. The scale is measured, not anecdotal. A large field study of hardware faults on consumer PCs and laptops, published in 2011 and run over eight months, counted four to 20 times as many crashes from CPU failure on overclocked machines, the range depending on the processor maker. Throttling earns its place in a glossary because it points at conditions. The other entries point at parts.
Where does a clock number actually come from?
The reference behind this page defines neither word. On a bench, the base clock is the reference frequency and the multiplier is the factor applied to it, and both are read from a board manual rather than from the article. Most of the surrounding vocabulary follows from that arithmetic. Many motherboards ship with extensive overclocking facilities built into the hardware and controlled from BIOS settings, and the multiplier is the usual lever, which is why locking exists. The article on overclocking records the speculation that manufacturers lock the multiplier to keep buyers from purchasing a cheaper grade and raising it, a measure marketed as consumer protection and criticized by buyers. Underneath sits binning. Parts from one manufacturing process are tested after manufacture and marked with a rating chosen by market needs; in a high-yield run, higher-rated parts can be sold as lower-rated, and a part's true maximum can exceed the highest rating sold. In the worst cases, running at that higher rating is more problematic.
Offset is a small word with a squared consequence
The word offset does not appear in the reference at all. On a bench it names a signed adjustment added to a reference value rather than substituted for it. The number itself carries no warning about weight, so the arithmetic has to. The record is explicit on shape: raising operating frequency raises thermal output roughly in line, while raising voltage raises thermal power quadratically. An offset applied to voltage therefore buys more heat than its size suggests. Stock cooling is designed for the heat a part makes at its rated speed, which is why overclocked builds add fans, larger heat sinks, heat pipes or water cooling, and why the desk keeps a plain account of what you actually risk when the heat has nowhere to go. Noise follows its own curve, roughly the fifth power of fan speed: halving the speed removes about 15 dB, and fan noise can reach 50 dB or higher depending on the speed, model and number of fans. An offset is something a bench hears as well as reads.
What does error correction do that a pass does not?
Error correction, as the bench uses the term, catches a memory error before anything downstream sees it; the pages read for this glossary name an EDAC subsystem in the kernel index and define nothing further. Its absence is what makes the darkest term in this glossary possible: silent data corruption, damage produced by undetected errors, which is how an unstable system can ruin data while appearing to work. A memory tester works the other side of the street. Memtest86+ is a free, stand-alone memory tester released under the GNU GPL v2, loadable by a PC BIOS, legacy or UEFI, or through a bootloader that supports the Linux boot protocol, and its own page describes it as more thorough than the memory check a BIOS performs. It is not the closed-source Memtest86, which has belonged to PassMark since 2013. Two cautions travel with the territory: the project's nightly builds are generated automatically and are fully untested, and the kernel's own documentation keeps an index of subsystems that lists an EDAC subsystem beside CPUFreq and watchdog support, which is as far as an index can take anyone.
Where the name on the address comes from
Borrowed, and the record is short. Overclockix first shipped in 2003 as version 3.2 on a Knoppix base, was inactive from 2005, was taken up again in 2011 on a Debian base, and its last published release is version 7.8.0 of June 1, 2015, as DistroWatch records it. Eight releases in all. This desk keeps the vocabulary and the methods, not the distribution; it serves no disk images, and the project's source is published on GitHub. One habit closes the loop. Before the next run, write down which of these eight words describes what the machine actually did. If the word is throttling, conditions get checked before parts get blamed. If none of the eight fits, write that down as well; an unnamed failure is still a measurement, and it is the one that sends a person to the documentation.
Entity record: "Overclocking", English Wikipedia, en.wikipedia.org. An encyclopedia article, not a procedure. It defines the practice as increasing the clock rate of a semiconductor device beyond its rated speed, then covers cooling methods, stability and reliability, the economics of rating parts, and factory overclocking. Its record reaches back to an unofficial modification of the UNIVAC I called the Overdrive, dated 1952, and forward to a CPU frequency record of 9,206.34 MHz set in May 2026, with the June 2006 announcement by IBM and the Georgia Institute of Technology of a silicon transistor switching rate above 500 GHz, cooled with liquid helium to 4.5 K, in between. Pages it points to include CPU-Z, Super PI and serial presence detect.