Idea for a new kind of GPU stability test

[Mathragh]

So today i was just stresstesting my GPU a bit after i've made some improvements to the cooling in my pc(added a piece of cartboard to prevent hotair from re-entering the GPU cooler )and it occured to me that with modern age GPU's that are all capable of doing things like OpenCL, stability testing could potentially be made a lot simpler.

CPU stability testing programs like IBT and Prime95 all have a way of totally stressing the CPU, and checking for errors at the same time, then reporting when an error has occurred, and thus that the CPU isn't stable.

Wouldnt it be alot simpler to have an OpenCL program that stresses the GPU to the fullest, and errorcheck at the same time aswell? That way the task of stability testing a GPU would be made a lot simpler; you could just let it run, and it will just report stability issues if errors are being made. Output will be in text only, no image rendering.

Now I dont know whether its even possible to program such a thing in OpenCL(or any other GPUcomputer languages), but I imagine that if it were possible, it wouldnt be that complicated to implement. It would also potentially make a great addition to our awesome inhouse GPU-Z program.

So to sum things up:
A program in OpenCL that will,
fully load the GPU by letting it do calculations(simple math), and
errorcheck all the calculated results to detect instabilities.
This al in a simple text based program(or built in some other program, like GPU-Z),
Maybe also with a sort of benchmark, but thats optional.


So, that was just me sharing an idea with you guys Feel free to comment/brainstorm on/improve upon/burn down/whatever on the idea

Happy Tweaking!

 

[EarthDog]

Doesnt Furmark Do that already? The Precision stress test... OCCT...MSI Afterburners stress test...

It may not be on OpenCL but it accomplished the same things already, no?

 

[Mathragh]

Doesnt Furmark Do that already? The Precision stress test... OCCT... It may not be on OpenCL but it accomplished the same things already, no?

Yeah, i'm using furmark aswell, but with stability test programs that render images, its not always easy to see artifacts, and requires more attention that a program like, prime95, which you can just run for a night.

 

[EarthDog]

I wouldnt run P95 overnight without checking it for a while first. Also, artifacts generally show up pretty darn quick for GPUs, especially in Furmark. So its a moot point for me... a good idea none the less.

 

[LAN_deRf_HA]

Furmark hasn't worked for me since my 260. It's completely unable to accurately error detect a modern gpu unless you're so unstable it can't even run aero. OCCT worked up until the 4xx series but now it's gimped by power limiting. The Crysis benchmark tool remains the most consistent and accurate way to stress test a modern gpu. That tells me a proper gpu test needs to be more than just a fuzzy render as clearly those don't fully stress the whole gpu to vram chain. Which is unfortunate because you have to buy Crysis to run it.

In theory the Metro 2033 benchmark tool could perform the same function, maybe even better since it's DX11, but I've only triggered an error with it once so far.