11
\$\begingroup\$

If you continuously run your graphics card at between 80°C and 90°C (176°F and 194°F) is it actually bad for the graphics card? I.e. does it reduce the life of the card? Can this be proven? Or is it just assumptions?

I understand that the safety shut off for GPUs is normally 90°C (194°F).

\$\endgroup\$
  • \$\begingroup\$ The 'safety shut off' depends a lot on where the temperature is measured and what process and maximum temperature the circuit was designed with. I remember a while back a certain generation of intel CPUs had a maximum rated temperature of 110°C, which got certain hardware-enthusiasts worried because they thought the chips would destroy themselves. Spoiler: They did not. \$\endgroup\$ – Joren Vaes Jun 20 at 11:10
  • 1
    \$\begingroup\$ I think this question is heavily related to this other question ic-product-lifetime-as-function-of-junction-temperature. Bottom line from that question is that every 15°C above room temperature halves an IC's life expectancy. So running a graphic card at 90°C compared to 80°C will reduce its lifetime by ~37% (so if life expectancy is 8 years with 80°C, it's ~5 years with 90°C instead) \$\endgroup\$ – Harry Svensson Jun 20 at 12:08
  • 1
    \$\begingroup\$ Arhennius Law is closer to 50% lower MTBF / 10’C rise but there are other factors to be considered for dielectrics where they start with a much lower MTBF like 1000 h @ 85h or 105h so I suspect they use 105’C rated caps or better. \$\endgroup\$ – Tony Stewart Sunnyskyguy EE75 Jun 20 at 18:17
21
\$\begingroup\$

Let us study the failure mechanisms, and see how they are affected by heat. It is very important to remember that just because a failure mechanism happens faster with temperature, the GPU will not necessarily fail faster! If a sub-component that lasts 100 years at room temperature only lasts 20 years if it is hot, but another sub-component only lasts 1 year to begin with (but is unaffected by heat), the lifespan of your product will hardly change with temperature.

I will ignore the cycling issue talked about by Simeon as this is not my expertise.

On the board-level, I can think of one main component that will 'break' with head: Electrolytic capacitors. These capacitors dry out, and it is well understood that they dry out faster when heat is applied. (tantalum capacitors also tend to have a shorter lifespan but I don't know how this changes with heat).

But what about the silicon?

Here, as I understand it, there are a few things that can cause failure. One of the main ones here is electromigration. In a circuit, the electrons going through bits of metal will actually physically move around atoms. This can get so bad that it will cause gaps in the conductors, which can then lead to failure.

This image gives a good illustration (from Tatiana Kozlova, Henny W. Zandbergen; In situ TEM observation of electromigration in Ni nanobridges):

enter image description here

This process increase exponentially with temperature, and thus indeed, the chip will last less time if temperature is higher and electromigration is the main cause of failure.

Anther mechanism is oxide breakdown, where inside the circuit the transistors will suffer gate-punch-through. This is also temperature dependent. However, voltage has a much bigger impact here.

There is also VT shift, either due to drift of dopants or due to hot-carrier-injection. Dopant drift increase with temperature (but unlikely to be an issue, esp. with digital circuits, as this is a very slow process). I am not sure about the temperature dependence of hot-carrier-injection, but I think again voltage is a much more important factor here.

But then there is an important question: How much does this decrease the lifespan? Knowing this, should you make sure that your graphics card stays cool all the time? My guess is no, unless an error was made at the design stage. Circuits are designed with these worst-case situations in mind, and made such that they will survive if they are pushed to the limits for the rated lifetime of the manufacturer. In the case of people overclocking circuits: The increase in voltage they often use to keep the circuit stable (as it can speed the circuits up a bit) will do far more harm than the temperature itself. In addition, that increase in voltage will lead to an increase in current, which will significantly speed up electromigration issues.

\$\endgroup\$
  • 2
    \$\begingroup\$ Those are some fantastic images, I have always wondered what electromigration would physically look like. \$\endgroup\$ – Cursorkeys Jun 20 at 12:14
9
\$\begingroup\$

Yes it is been proven that heat degrades electrical components. Metals expand when they heat up, solder (used for electrical circuit connections) is a metal alloy so it will expand when heated up. Constant heating and cooling will cause the joints to constantly expand and contract which can lead to cracking and eventually failure of the joint.

                                                      Graph of Failure rate vs Temperature

The graph above shows how Arrhenius'Law gives a correlation between an increase in heat and semiconductor failure. This paper details the effects of heat on electronic components. It deals more with things at the electron level, which is a bit outside my scope of knowledge

\$\endgroup\$
  • 1
    \$\begingroup\$ I can believe that cycling is bad, due to as you say expansion and contraction, but is there a problem with running at a high load, and thus a high temperature all the time? \$\endgroup\$ – Colin Jun 20 at 8:05
  • \$\begingroup\$ I am a IC designer, so I have little knowledge about board-level failure modes, but in all my time repairing things (as a hobby) I have yet to come across a failure due to expansion cycling, so I have to question how significant it is compared to other mechanisms. \$\endgroup\$ – Joren Vaes Jun 20 at 11:04
  • 1
    \$\begingroup\$ @Colin no such thing as "high load all the time"; unless you're just e.g. mining bitcoin on your GPU, there will be seconds where there is more load than others. With cooling having to be pretty potent on GPUs, this already leads to the mentioned problems. See: XBox ring of death. \$\endgroup\$ – Marcus Müller Jun 20 at 11:39
  • \$\begingroup\$ @MarcusMüller there absolutely is. And It doesn't matter that the load is not absolutely constant. For cycling the delta temp is important. A card that runs 99% of the time at 95-100% load (i.e. compute) within the designed temps, would be far less susceptible to the hypothetical cycling damage, than the same card oscillating between 0% and 100% wildly 50% of the if (i.e. games). \$\endgroup\$ – Dan M. Jun 20 at 14:38
6
\$\begingroup\$

The relationship between the increase in the junction temperature of a semiconductor and the reduction of its MTBF (Mean Time Between Failure) is well understood.

This technical note from Micron talks about this

In practice, the failure rate will increase exponentially once the junction temperature approaches and exceeds ~125˚C, so if you are operating well below that temperature small increments may not be that critical.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.