What can a memory test, a SMART attribute table, and a temperature or voltage sensor each prove about a machine, and where does each one stop?
Start with the glass
A memory test exercises cells under load and reports the errors it saw. A SMART table reports what a drive noticed about itself. A sensor panel reports voltage and temperature at a few points on a board. Each is a different instrument answering a different question, and The Burn-In Desk, a hardware diagnostics desk built on the principle that every tool has a blind spot worth naming, is the right companion for learning which instrument saw what.
Separate the variables
A memory test writes patterns, reads them back, and counts mismatches. A clean pass under one pattern and one duration says the tested cells behaved under those conditions. It says less about the access pattern a different workload creates, about errors that appear only when the machine is fully heat soaked, or about a fault that needs two specific slots populated. A failure localizes suspicion to the memory subsystem; isolating a single module still means testing one stick at a time, because a dual channel fault is often a pairing or a slot, not a bad part.
Make the note useful
SMART attributes are counters the drive keeps about itself: reallocated sectors, pending sectors, interface errors, power on hours, temperature history. The raw values matter more than the normalized scores, because each vendor scales them differently. For reading them, the smartmontools reference implementation documents the attribute families most tools wrap, and even a clean table only says the drive has not recorded certain failure classes yet. A counter that moved since the last check is information; a counter that never moved is not a promise.
Respect the limits
Temperature and voltage readings carry their own boundaries. A motherboard sensor reports where it sits, not the silicon you care about; a socket temperature and an internal junction can differ by tens of degrees under load. Voltage rails read through a monitoring chip are not calibrated instruments, so the trend and the shape of a dip under load mean more than the digit. Two tools reporting different numbers are usually sampling different registers, not disagreeing about physics.
A small next step
The method that survives is a log, not a heroic test session. Record which tool ran, which version, which pattern or attribute set, how long, and what the machine was doing. When results conflict, remove variables instead of adding tests: a bootable toolkit takes the installed system out of the diagnosis, one memory module takes the pairing out, a cold start takes the heat soak out. A machine that fails after every clean test is still telling you the boundary, not refusing to answer. The same discipline serves the brew bench, where a hydrometer reading is one instrument with one boundary.
How long should a memory test run before it means anything?
Long enough to reach the failure mode you fear. Most pattern errors surface in the first passes, while thermal and marginal timing faults may need the machine fully warm, so a ten minute pass and an overnight run answer different questions. The practical rule is to run until the conditions that produce your symptom have occurred: a machine that crashes in hour four of a render needs a test that sees hour four, not a summary of minute ten. If a suspect module passes overnight, move it to another slot before trusting the result, because slot and channel are part of the diagnosis.
Which SMART attributes deserve attention first?
The counters that describe permanent loss: reallocated sector count, current pending sectors, and uncorrectable errors on spinning drives, plus the percentage used and spare capacity figures on solid state models. Interface CRC errors often point at a cable rather than the drive. What matters is the delta: an attribute that stays at zero for years is boring, and a count that grew between two checks is a schedule for backup and replacement, whatever the normalized score says.
Why do two tools report different temperatures?
Because they are reading different silicon. A monitoring chip exposes several inputs at once, tools pick different ones, and each sensor has an offset the software may or may not apply. Package temperature, per core sensors, and the socket thermistor describe different physical points, so quoting one number without naming the source is where most arguments about overheating go wrong. Trust the trend inside one tool, not the absolute difference between two.
What is a diagnostic worth when the machine still fails?
It narrows the search instead of ending it. A log of clean memory passes, a stable SMART table, and sane voltage traces removes whole families of causes, which is the real deliverable: the remaining suspects are fewer and cheaper to test. The honest outcome of diagnostics is a shorter list, and a record that lets the next fault be compared against a baseline rather than a memory.
Context before numbers: What can a memory test, a SMART attribute table, and a temperature or voltage sensor each prove about a machine, and where does each one stop?
