How to Read GPU Benchmark & Stress-Test Results: Scores, Clocks, Temperatures, Power & Fault Signals Explained
重點摘要
Benchmark scores and stress tests answer different questions, and neither number means much alone. This guide explains how to read scores, clock curves, core, hotspot and memory temperatures, power behaviour and fan response together, separate throttling from faults, and decide when a result warrants inspection rather than another test run.
What This Guide Adds: Reading Results, Not Just Running Tests
Earlier articles on this site, including guides on FurMark, driver resets, undervolting and temperature differences, use benchmarks and stress tests as diagnostic steps: run a test, watch whether anything goes wrong, and move on to the next step. That approach is useful, but it leaves the most important question unanswered. Once a test has finished, what do the numbers actually mean?
This article works at a different layer. It does not teach how to install or launch a test, and it does not repeat a step-by-step diagnostic procedure. It teaches interpretation: which numbers carry meaning, which numbers mislead, how to read several signals together instead of one number in isolation, and how two properly compared runs can support a conclusion that a single run never can. If you have ever finished a benchmark, looked at a score, and still not known whether your card was healthy, this is the missing layer.
Two Different Tools That Answer Different Questions
The first interpretation error is using the wrong tool for the question. A synthetic benchmark and a stress test look similar on screen — both load the card and report numbers — but they measure different things, and swapping their jobs produces confident wrong answers.
A benchmark, such as 3DMark Time Spy or Steel Nomad, or Unigine Superposition, measures completion speed under a fixed combination of loads. It answers the question: how fast does this system finish this particular workload? The score is a product of the whole system, not of the card alone. Driver version, background load, power plan, cooling state, PCIe bandwidth, and CPU bottlenecks all feed into the same number. Two identical cards in different systems can legitimately produce different scores, and neither score proves that one card is faulty.
A stress test, such as FurMark or the GPU test in OCCT, measures stability and cooling behaviour under sustained full load. It answers a different question: does the card hold together, without crashing, artefacts, or runaway temperatures, when pushed hard for a period of time? A stress test is not a performance measurement. Passing one says nothing precise about how fast the card is, and failing one does not by itself say the card is slow.
FurMark deserves a specific warning here. Its load pattern is more extreme and more uniform than a typical game: simple geometry, very high power density, and little of the variation that real game engines produce. Temperatures and power draw recorded under FurMark therefore cannot be carried across to gaming directly. Judging whether a card will overheat in games from its FurMark temperature will overstate the problem. Conversely, surviving FurMark does not guarantee stability in a ray-traced title or a high-memory game, because those workloads stress different parts of the card. The two kinds of load overlap only partly.
| Tool | What it measures | What it cannot prove |
|---|---|---|
| Synthetic benchmark (e.g. 3DMark, Superposition) | Completion speed under a fixed load combination, for the whole system | Hardware health, or stability under sustained load |
| Stress test (e.g. FurMark, OCCT GPU test) | Stability and cooling behaviour under sustained full load | Performance, or stability under different load types such as ray tracing |
| In-game measurement | Behaviour under one real workload | General stability, or comparability with other systems |
The Five Signal Groups, Read Together
No single number diagnoses a card. A defensible reading combines five signal groups, and each group is described below in the order to check them.
1. Scores: Only Compare Like With Like
A score is only meaningful against a valid reference. The two valid references are your own machine's earlier score under the same settings, and a published reference range for the same card model in a comparable class of system. A score compared across generations, across very different systems, or against a single number quoted for a different configuration proves nothing.
When a score falls, work through the mundane causes before suspecting hardware ageing: has the driver version changed, has the power plan changed, is something running in the background consuming resources, has the case temperature risen, is the card now sharing PCIe bandwidth it previously had to itself? Until those are excluded, a lower score is an unexplained change, not evidence of a fault. A score that returns to its previous level after one of these is corrected was never a hardware problem at all.
2. Clock Curves: Shape Matters More Than Peak
The headline clock speed printed on a box, and the peak value shown in a monitoring overlay, are the least informative numbers in the whole report. What matters is the shape of the clock curve across the whole load period: does the clock rise and hold steady, does it slide downwards gradually, or does it collapse suddenly and recover?
A clock that starts high and then slides steadily downwards is the classic shape of a thermal or power limit engaging. A clock that drops off a cliff and then jumps back up points in a different direction: a driver reset, a power protection event, or instability. A clock that never rises far in the first place, while utilisation stays low, usually means the card is not actually being loaded — the bottleneck is elsewhere, or the test itself is not feeding the card enough work. Clock readings must always be read next to temperature readings; a clock curve without its temperature curve is guesswork.
3. Temperatures: Three Numbers, Not One
There is no single GPU temperature. Three readings tell three different stories, and collapsing them into one number hides exactly the information that separates one fault from another. Core temperature describes the main graphics processor. Hotspot temperature describes the hottest measured point on the die, which can run well above the core reading and is often the first reading to reveal a mounting or contact problem. Memory junction temperature describes the memory chips, which have their own thermal behaviour and their own limits.
What to look for is the shape of each curve. A temperature that rises to a plateau and holds steady there describes a cooling system reaching equilibrium: heat in equals heat out. A temperature that climbs without levelling off describes a cooling system that cannot keep up. Absolute values must be judged per model, because architectures and cooler designs differ widely; this article deliberately gives no universal "normal temperature" table, because any such table would be wrong for many cards. The valid comparison is your own card's readings before and after a change, with all three temperatures read together.
4. Power: Limit Behaviour Versus Fault Behaviour
Power draw must be read against the card's power limit and its design power. Power sitting at the limit with clocks held below their peak is power-limit behaviour, not a fault: the card is doing exactly what its power management is designed to do. Power that is unusually low while utilisation is also low points away from the card entirely, towards a bottleneck elsewhere — the processor, PCIe bandwidth, driver settings, or a test that is not loading the card fully.
The pattern that does warrant attention is power collapsing suddenly to near zero together with a black screen, a driver reset, or a full system restart. That pattern belongs in power-supply diagnosis, not in test-software blame. A single power reading, without the clock and temperature readings taken at the same moment, cannot distinguish these cases.
5. Fan Response: Does Speed Follow Temperature?
Fan speed should rise as temperature rises. When temperature climbs but fan speed does not respond, the problem sits in fan control or fan power delivery. When fan speed rises normally but temperature climbs anyway, the problem sits elsewhere: in cooler contact, in case airflow, or in a heat path that the fan cannot fix by spinning faster. Both patterns look like "overheating" in a one-line summary, and they lead to completely different next steps. Judging a fan by noise alone, or by speed alone without the temperature curve beside it, repeats the same single-number error.
Four Causes of Throttling, and How to Tell Them Apart
Throttling is often treated as a synonym for failure. It is not. Throttling is a protection mechanism: the card deliberately slows itself to stay within a limit. The diagnostic job is to identify which limit is engaging, because each cause has different implications and different remedies.
| Cause | Typical signature | Test continues? |
|---|---|---|
| Temperature limit | Temperature high, clock sliding down gradually | Yes, at reduced speed |
| Power limit | Power at its limit, temperature not necessarily high, clock held below peak | Yes, at reduced speed |
| Voltage or supply instability | Clock dropping suddenly, flicker, driver reset, or system restart | Often interrupted |
| Driver or settings change | Appears after a driver update or settings change, disappears when reverted | Yes, behaviour changed |
The decisive question is what else happens alongside the slowdown. Throttling plus artefacts, crashes, black screens, or driver resets is a different situation from throttling alone. Throttling alone, with a clean image and a completed test, is the card protecting itself. Treating every throttled run as a fault leads to unnecessary part replacement; treating every fault as "just throttling" leads to missed diagnosis. The accompanying symptoms draw the line.
Fault Signals and Normal Phenomena: The Boundary
The following boundary is the practical core of this guide. Results on the left warrant further investigation; results on the right do not, on their own, indicate a fault.
| Observation | Reading |
|---|---|
| Artefacts or broken rendering at stock settings, persisting across driver versions | Fault signal: investigate further |
| Black screen, driver reset, or full system restart during a test | Fault signal: investigate further |
| Fan not spinning while temperature climbs | Fault signal: investigate further |
| Same test repeated, results getting worse each time | Fault signal: a deteriorating trend |
| Single score lower than an online reference | Normal on its own: references differ by system, driver, and cooling |
| FurMark temperature higher than gaming temperature | Normal: FurMark load is more extreme and more uniform |
| Clock held back by the power limit | Normal: power management working as designed |
| Cold first run scoring differently from a warm later run | Normal: cooling state changes results |
| Different scores across driver versions | Normal: drivers change performance behaviour |
Three boundary cases deserve care. First, artefacts that appear only under overclocked or undervolted settings and vanish at stock settings indicate an unstable setting, not faulty hardware. Second, artefacts that appear at stock settings but only in one specific test, while other tests and games are clean, should first raise suspicion about that test, its driver interaction, or that specific load combination; cross-check with a different test before concluding anything about the card. Third, intermittent faults can hide completely inside a short test: a normal test result means only that this test, at this time, did not trigger the problem. It does not exclude a hardware fault.
A Repeatable Reading Routine
A single run proves very little, because a cold first run and a warmed third run can differ substantially as the cooler absorbs heat. A defensible reading needs repetition under controlled conditions:
- Fix the conditions: same machine, same driver version, same power plan, same test, same settings, and no background load.
- Run the same test at least twice, and note the trend across runs rather than the single best or worst number.
- Record all five signal groups together — score, clock curve, three temperatures, power, and fan response — not whichever number the overlay shows most prominently.
- Change one variable at a time. Change the driver, or the power plan, or the case airflow — never several at once — and re-run under otherwise identical conditions.
- Compare against your own previous runs first, and only then against published reference ranges for the same model.
If a result changes when one variable changes, that variable is part of the explanation. If a result persists across driver versions, settings, and cooling states, it is more likely to be a property of the card itself. That discipline — one variable at a time, repeated runs, all signals together — is what turns a benchmark from a number into evidence.
Five Misconceptions, Corrected
"A low score means the card is ageing or failing." Not until driver version, power plan, background load, cooling state, and CPU bottlenecks have been excluded. Most unexplained score drops have mundane causes.
"Surviving FurMark proves the card is stable." FurMark's load is narrow. Passing it does not prove stability under ray tracing or high-memory game loads, and failing it does not prove games will fail either. Real-load verification is the necessary second step.
"One temperature number is enough." Core, hotspot, and memory junction temperatures can tell different stories, and a single reading hides the difference between a contact problem and an airflow problem.
"Benchmark scores can rank cards across machines." Scores from different systems, drivers, and cooling setups are not directly comparable. Buying decisions need same-class comparisons, not single numbers.
"A test result of PASS means there is no hardware fault." Intermittent faults, such as board-level connection problems that appear only under certain thermal or mechanical conditions, can be entirely absent from a short test. A clean result narrows the possibilities; it does not close the case.
When to Stop a Test Immediately
A stress test is a diagnostic tool, not an endurance contest. Stop immediately if artefacts appear, if there is an unusual smell, if a fan develops an abnormal noise, or if temperature climbs without levelling off. Continuing to "see what happens" after such a signal only increases the risk of further damage. Note what was on screen, what the readings were at the time, and which test was running; that record is far more useful for diagnosis than a completed run achieved by pushing past a warning.
When Inspection Is Warranted, and What to Prepare
Inspection is warranted when fault signals persist at stock settings across driver versions, when repeated runs show a deteriorating trend, when a fan fails to respond to rising temperature, or when black screens, driver resets, or restarts recur during testing. Before bringing the card in, record the test used, the settings, the driver version, the readings at the time of the problem, and whether the problem appears in more than one test. That record lets an inspection start from evidence rather than from scratch.
HK GPU Fix carries out in-shop inspection at Room 1209, Dali Plaza, Shek Mun, telephone 2186 8010, WhatsApp 6558 6806. Published inspection and repair pricing sits in the three-tier range of HK$400–2,800, with the actual figure quoted after inspection. Bring the card with its power cables and a note of the readings described above; a clear symptom record shortens diagnosis considerably.
FAQ
Can a single benchmark score prove my graphics card is faulty?
No. A score is a whole-system measurement influenced by driver version, background load, power plan, cooling state, PCIe bandwidth, and CPU bottlenecks. A single score can only show that something differs from a reference; it takes repeated runs under controlled conditions, with the mundane causes excluded, before a score points towards the card itself.
Why is my FurMark temperature higher than my gaming temperature?
Because FurMark's load is more extreme and more uniform than a typical game, with higher power density. FurMark temperatures should not be used to judge gaming temperatures directly. Use FurMark for stability and cooling behaviour under sustained load, and use real game loads for gaming temperature expectations.
My clock speed drops during a long test. Is the card failing?
Not necessarily. A gradual clock decline alongside high temperature or power at its limit is throttling, which is a protection mechanism, not a fault. The warning signs are what accompany the slowdown: artefacts, black screens, driver resets, or restarts. Throttling alone, with a clean image and a completed test, is the card protecting itself.
Do I need to watch core, hotspot, and memory temperatures, or is one enough?
All three. Core temperature, hotspot temperature, and memory junction temperature can tell different stories: a contact problem may show up first in the hotspot reading, while a memory cooling problem may show up only in the memory reading. Reading one number alone can hide exactly the distinction that determines the next step.
My test result is normal, but the problem still happens sometimes. Does that rule out hardware?
No. Intermittent faults can be entirely absent from a short test. A normal result means only that this test, at this time, did not trigger the problem. If symptoms recur outside testing, keep a record of when they occur, under what load, and at what readings, and use that record as the basis for inspection.
Should I compare my score with scores posted online for the same card?
Only as a rough range, and only for the same model in a comparable class of system. Scores vary with driver version, cooling, power plan, background load, and the rest of the system. The strongest comparison is always your own machine's earlier score under the same settings; the second strongest is a published range for the same model, not a single number from a different system.