3D wave simulator overheats my HD 5870

[Voxilla]

Here another DX11 demo, this time a 3D volume wave simulator.

Unfortunately after running for a couple of minutes my HD 5870 GPU seems to overheat. Temperature rises to 87 degrees Celsius. At this point either the simulation gets corrupt at certain places in the volume.
It gets worse when a vertical stripe pattern appears all over the screen.
The only way out seems to be a hard reset of the system.

Am I the only one with the phenomena ?

 

[AlexV]

Hrm, that sounds familiar...try monitoring your Engine/Mem clocks in the background via GPU-Z, see what happens when you reach the area around 90. It should downclock to 600E/900M normally...it also shouldn't get that high. Is the fan spinning up properly? The first sample I had experienced some trouble on that front.

 

[Voxilla]

With temperature at 87 degrees Celsius, the fan runs at 44%.
GPU and memory remain at 850 / 1200 Mhz.

 

[AlexV]

Hmm, and at 44% it doesn't manage temperatures properly? That's very odd...I've played with your demo (very nice, by the way!), and it is indeed a scorcher, but temperatures topped out at 85C here, and the fan always kicked in to about 37% whenever needed, keeping them there. Very odd...

 

[Voxilla]

Maybe my GPU is just marginally Ok, and just not Ok on unusual high load.
It overclocks to only 860 Mhz. There is little chance of getting it replaced I suppose.

 

[Albuquerque]

Maybe my GPU is just marginally Ok, and just not Ok on unusual high load.
It overclocks to only 860 Mhz. There is little chance of getting it replaced I suppose.

I don't think you could get it replaced based on overclocking capacity. But if you can show a workload in which the card should perform normally where it fails, then I would believe that to be sufficient argument for replacement.

 

[AlexV]

Maybe my GPU is just marginally Ok, and just not Ok on unusual high load.
It overclocks to only 860 Mhz. There is little chance of getting it replaced I suppose.

I'm not sure that there's anything wrong with your card...I think you might have coded the 5800's Furmark

 

[Voxilla]

I don't think you could get it replaced based on overclocking capacity. But if you can show a workload in which the card should perform normally where it fails, then I would believe that to be sufficient argument for replacement.

No, for me it does not have to overclock at all, I just said this to indicate that my particular GPU might be marginal.

My wave simulator fails to perform normally, I think this should not happen.
If a similar program would fail running on a CPU, there is little doubt the CPU would be declared faulty.

 

[Florin]

No, for me it does not have to overclock at all, I just said this to indicate that my particular GPU might be marginal.

My wave simulator fails to perform normally, I think this should not happen.
If a similar program would fail running on a CPU, there is little doubt the CPU would be declared faulty.

Sadly there is still some doubt. Just look at Furmark discussions here. There's actually quite a lot of people out there who get rather defensive about this - they would just claim that your program is a 'power virus' or some such nonsense, that it creates an unrealistic load and should be ignored.

 

[fellix]

I've asked a friend to run the program on his overclocked 5870 (1000MHz core, 1.25V), and with default fan speeds (auto) the GPU temperature tops 90°C.

 

[hoom]

Furmark is supposed to run perfectly fine & safely on R8xx because they put in the new hardware monitoring thingie that downclocks automatically when it gets too hot & then back to full speed when its sufficiently cooled again.

I'd guess a test for if the problem here is heat would be to set fan to an excessively high rate (75%/100%) when running it.

 

[Psycho]

My 5770 goes to 65 degrees and 45% fanspeed - while furmark tops around 69 degrees. So it's no so much a power virus on the "half 5870".

But, does the 5000 series throttling have anything to do with heat or isn't it just the vrm load?

 

[hoom]

Ok, yeah I think it does both

 

[Voxilla]

I might have found a satisfactory solution without having to replace the card,
by adding another fan to the case.
This has a noticeable effect on maximum GPU temperature and fan speed.

Temperature and fan speed are now 87 degrees / 40% down from 88 / 44.
The system seems to remain stable now running the wave simulation.

Furmark still is able to generate a good deal more heat.
Before 90 degrees at 53 % fan speed (if the system would not crash)
Now 88 degrees at 47 % fan speed.

So having a very good ventilated case is very important with the HD 5870
The cause of this, I think, is with the design of heat exhaust of this card, which for 30 % ends up back in the case.

 

[Voxilla]

Furmark is supposed to run perfectly fine & safely on R8xx because they put in the new hardware monitoring thingie that downclocks automatically when it gets too hot & then back to full speed when its sufficiently cooled again.

I didn't see this safety mechanism kick in action, so I'm very doubtful that it is there or enabled. Maybe it kicks in action for temperatures higher than 90 degrees, but these are definitely too high for my GPU.

In the Catalyst panel there should be an option, I think, to force the fan like for example 10% higher rpm than the automatic value. Now it can only be forced at a fixed high speed, but this is not acceptable because of the constant noise.

 

[CarstenS]

I have both a reasonably well ventilated case and a HD 5870 and no problems running your psychedelic LSD-simulator (*SCNR*). You might want and try MSIs Afterburner-Tool. In it you can set a user-defined fanspeed curve (plus over- and undervolt your card) which you can activate and deactivate with one click. Very nice!

 

[Voxilla]

About the demo, this is actually the slow version.

I have another version running twice as fast by storing the volumes as 4 channel floats instead of single channel.
It get's a little messy then to update the sources as you can not write to single channels of an unordered access texture resource, sigh ...
Also the rendering then can not make use of the hardware texture filtering.
There seems no way to type cast a 4 channel view in a 1 channel view....
Reading from a RWTexture3D is also impossible if it is rgba, can be only single channel.

I'm experimenting with more interesting ways to visualize the waves by means of for example marching cubes or raycasting.
Now that I think about it I probably should do something about the wave sources to make the simulation run similar on slower hardware...

 

[oscarbg]

Voxilla relase the code in case are you getting 250fps..
I want to see your update RW buffers code..
Thanks

 

[Freak'n Big Panda]

I'm so impressed with these demos! How long have you been coding for voxilla? I've just started with 3d programming (going through OpenGL tutorials lol) and I have to say that it can be quite tedious at times. Enjoyable though, once you get the ball rolling at least!

 

[rpg.314]

OT

@panda: you'll want to avoid the fixed function calls in case you are just beginning with ogl.