GPU is too hot, can I turn it down?

[Mispey]

I was glad to notice v7 works with my GPU automatically and I'm REALLY Folding now.

But I'm kind of uncomfortable with the steady 65C my GPU holds while folding. When I turn the fans on the GPU up to 100% it only drops to 60C, and it's loud. Either case here is not really acceptable for me.

What can I do to lower the load on my GPU? As far as I can tell there is only an option to lower CPU usage. Otherwise, I need to stop folding with my GPU.

[Ripper36]

What GPUs are you using? I WISH mine were only 65C - doesn't seem hot to me.

If you must, you can turn the clock speed down and this will reduce temp. Use the control panel that comes with your GPU or something like MSI Afterburner

[Zagen30]

Yeah, 60-65C is not very hot. GPUs are designed to handle temps up to around 100C without failing. I'm not sure I'd want my GPUs running consistently in the 90's, but 60's and 70's should be safe.

[P5-133XL]

Different video cards with different cooling systems vary extensively in how well they cool and how loud they are. Some fans can be outright annoying at 100% and you will shorten their lifespan by running them at full speed.

Not trying to counter your comfort zone but 60C-65C is actually cool for video chips. Modern GPU chips are designed to tolerate significantly hotter temps. I'm talking 100+C before they start self-protecting themselves from heat issues by clocking themselves down. So you have lots of head-room.

Probably the easiest way to deal with heat is to download yourself an OC'ing utility like EVGA's precision and then down-clock the card, adjust the fans, or if you have a video card capable drop the voltage somewhat. However, you can't drop the voltage too much or you'll start seeing errors (NAN's or EUE's) in your folding results and that is counter productive.

It may take a bit of fiddling to get it just right but I assure you that you can maintain good folding productivity, reasonable temps, and low noise.

[Mispey]

Thank you to all of the above responses. I for some reason had it in my head that those were rather warm temps for a GPU. I guess I don't really feel the need to do anything! Thanks.

[Dark Pulse]

Nope, 60-65C isn't even hot for a CPU, for that matter.

On my GPU, they tend to run about 85C while folding.

Really, 90+ is when it's starting to get "too hot." Anything under that and you're pretty much fine.

[iceman1992]

Mine runs at 75C. Just curious, what video card do you have??

[kiore]

Too hot is relative, if it is so hot it is causing noise issues for example then you should wind it down. But for many cards that is not hot at all, my old GTX 260s frequently sat on the high 80s low 90s for months on end while my GTX 560tis rarely go much over 70c unless the room is very hot.
My previous 8800gts 9800gt's and GTX275s all ran well over 70c, again 24/7 for years and the only fan failure I had was caused by me dropping a case screw on the spinning fan smashing off half the blades, even then it still worked if I disabled the fan and just let it run passively.
Some cpus run at high temperatures too, but not all, my old Phenom II is getting pretty hot at 50c.

[Dark Pulse]

Mine runs at 75C. Just curious, what video card do you have??

A GeForce GTX 690.

[bruce]

The GPU (and the CPU) will always run at 100% if it can until something stops it (typically running out of things to do) and then wait for the next instructions are ready to be processed. Managing throughput can really only be handled by withholding work batches, leading to lower averages.

Some time back, the Pande Group did some work with Environment Variables to allow the GPU utilization to be reduced. That tends to also reduce the temperature. Using that method of GPU adjustment has not been particularly effective. The fundamental problem is that running at temperature X or Y probably puts less stress on the GPU than cycling between X and Y. Bottom line: The hardware manufacturer is responsible for managing temperature, not some programmer writing software.

Since your temperature is already "cool enough" this doesn't really apply to you, but it's something to consider.

[jrweiss]

Some time back, the Pande Group did some work with Environment Variables to allow the GPU utilization to be reduced. That tends to also reduce the temperature. Using that method of GPU adjustment has not been particularly effective. The fundamental problem is that running at temperature X or Y probably puts less stress on the GPU than cycling between X and Y. Bottom line: The hardware manufacturer is responsible for managing temperature, not some programmer writing software.

First, I doubt that rapid cycling of the GPU load (in context of pausing 5 or 10 cycles every 100, or 50-100 in every 1000, on a multi-MHz GPU) will cause significant, instantaneous temperature variations. Since all my GPUs have been ATI/AMDs, and that parameter was only implemented for nVidia GPUs, I could not test it directly. Constantly changing load is inherent in any CPU or GPU implementation, regardless of what the average load is.

Second, IMO the GPU mfgr can't be held responsible for all implementations of his product. I don't know what the specified load cycle is that's considered when GPU thermal management is designed. Is the GPU mfgr responsible for case design and end-user implementation of cooling flow in his box? Are GPUs all designed for constant 100% load at nominal warm room temps (85F/30C)? [For a reference point, the Thermal/Mechanical Design Guide for the Q9xxx CPUs appear to be based on a max air temp of 40C at the CPU fan outlet/heat sink interface at full CPU load of 95W.]

I think the GPU utilization parameter is just as viable as the CPU load parameter for thermal management. Though the OP may not need it, it may be useful in systems with marginal cooling.

[bruce]

You're asking exactly the right questions. Under what conditions are GPUs tested by the manufacturer and what good is their warranty if they fail?

They have described a reference design. I don't know how to find out the definition of the reference loading profile except to ask them for an official definition.

CPU manufacturers do communicate reference cooling requirements to computer manufacturers which include heatsink requirements, case airflow considerations, and heating from other components in the vicinity. The fact that "unofficial" computer manufacturers don't know those recommendations plus they overclock / under ventilate / use alternate heat sinks / etc. does NOT place constrains the software programmer.

In my field (aerospace) every aircraft is certified for a certain number of operations based on things like is it a regional (short-range) jet that's going to see more takeoffs and landings per day than the jumbo jets (long-range). Both of those numbers are multiplied by the life of the aircraft with a number of strong assumptions making the numbers more conservative. Repair/servising requirements are also specified by the component manufacturer AND COMMUNICATED TO THE FINAL CUSTOMER in both the service manual and the warranty information. Personally, I have never read the service manual or even the detailed warranty information for GPUs but that doesn't mean they don't exist. Of course the FAA gets deeply involved in certifying aircraft and having you say "My GPU has been running at 95C for two years" doesn't carry as much weight as the crash investigator who has to decide if it was designed wrong, serviced wrong, manufactured wrong, piloted wrong, overloaded, or any number of other possible conclusions and they write corrective actions to assign to everybody involved.

GPU manufacturers do test GPU in a hot environment with what some engineer decided was the maximum workload that he assumed could ever be applied to the GPU. That number was probably increased when they started advertising that the GPU could be used as a stream processor, not just to display graphics -- unless it was already 100% -- and he had to consider the fan profile, reference design, some limiting value for overclocking, etc.

[k1wi]

It was mentioned in the first reply to the thread that downclocking core speed of the GPU is a method of reducing temperatures. I see this as a manual way for users to do what the latest generation GPUs should be increasingly doing now, which is controlling performance/heat through dynamic frequency scaling. The one issue that that creates is that it lowers the global performance of the GPU. Application utilisation limits & priority control would be a logical hypothetical solution but is hindered by the absence of the sophisticated hardware/firmware management required. These are all largely a manufacturer's 'problem' but one that does not end up with something that looks good on a marketing slide...

I know that in the past GPU manufacturers consider applications that used as much power as FAH (Furmark) to be 'power viruses' and I think AMD(?) even went so far as to detect when it was running and crimp power output. I put that down to them not having anticipated the rise of GPGPU & GPU compute. I don't think that the power virus claim stacks up any more as compute is increasingly becoming a mainstream activity of GPUs, even games (Civ5) is utilising compute functions.

Under what conditions are GPUs tested by the manufacturer and what good is their warranty if they fail?

I think there is even more than just if their warranty fails - also important is how long should a GPU last & what directly influences GPU life? Part of the clock rating is surely based on the MTBF rate & a target probability of failure during the warranty period. However, most consumers expect their hardware to last longer than the warranty period.

[bruce]

Under what conditions are GPUs tested by the manufacturer and what good is their warranty if they fail?

I think there is even more than just if their warranty fails - also important is how long should a GPU last & what directly influences GPU life? Part of the clock rating is surely based on the MTBF rate & a target probability of failure during the warranty period. However, most consumers expect their hardware to last longer than the warranty period.

True, but remember that the goal of a business is to make money. One of the goals of a manufacturers test is to control the costs of warranty replacements. That's quantifiable and serves as a minimum requirement.

Another is to maintain a good corporate image supporting the sale of future hardware. They may choose to provide a warranty replacement even after the warranty has expired ... or not. It was pretty well known that some early GPU chips were subject to cracking and that was partially fixed by the manufacturing-defect replacement clause and partly fixed by a driver change that reduced maximum heat (and FAH performance) and some of it probably wasn't fixed.

[prjindigo]

Changing the core-clock speed on a GPU will not change the sink temperature.
The only way to change the sink temps without using an OC utility is to correct your case ventilation.
I'd say about 100% of the cases I've seen have insufficient ventilation.

If you seriously want an overclock utility just for the purpose of fan speed control I'd advise MSI's Afterburner.
It has temp curve control that can support independent settings on multiple cards.
Keep in mind to never tell a card to run at 100%, some programmers don't understand that "256" doesn't exist.