PC Graphics

[tetedeiench]

Hi Everyone,

I'm the OCCT Dev ( https://www.ocbase.com ).

I was recently warned that there was an issue with AMD GPUs and my VRAM test.

Testing has shown that using any driver above V 20.4.2 result in my test being unable to allocate the memory - and then crashing.

I tried to contact AMD to no avail yet - as i do not own an AMD GPU, it's rather hard for me to replicate and debug this issue.

I updated my website to make it clear there's an issue - it's either on my side or on the driver side (at that stage, odds are on the driver side, but you never know ).

Sorry for the weird behaviour,

Tetedeiench

[elstaci]

just curious but, in layman terms, How does OCCT actually check a GPU's Video RAM?

I know that most diagnostic software checks regular system RAM by inputting a string of simple to complex numbers or characters and then checking the output to see if it is exactly the same as the input or if it is different, no matter how small the difference is. Any difference would be considered to be an "error" thus possibly defective.

I imagine a GPU Video RAM is different as far as checking the actual RAM since it is integrated on the Main board.

[tetedeiench]

It's actually pretty similar - When using OpenCL, you can allocate and use memory *almost* as if you're using system memory.

So testing it is indeed writing to it and checking the results. Doing so isn't as easy as on the System RAM because of the heavy caching that's occuring, and the heavy parrallelism that goes on there.

When my VRAM test does a pass, it does more than 10 different test patterns, some random, some being "key patterns" like walking ones, malking zeros... etc.

[elstaci]

You mentioned that you found the problem causing the issue and need Users with AMD GPUs to test it out.

Here is the developer's email to me concerning the fix ( I have a Nvidia GPU card):

On 9/16/2020 11:54 AM, Tetedeiench wrote:

I actually nailed it - AMD drivers are reporting a value that doesn't make any sense for memory allocation.

I put a cap on it and it seems to work fine again.

Are you affected by this issue ? If so, i think i'll have a working version ready for testing in a few hours - do you mind testing it then ?

Thanks,

Adrien         

Plus this next email from the developer:

It's actually a cap on the memory chunk allocation size ( i'm allocating several chunks until things are working again).

It seems to be running again - and without that crazy value, AMD's driver is behaving normally and much more nicely (not crashing at all, telling you when you overallocate, no crashes, etc).

Seems i was a tad too fast in my judgement of their driver, as the wrong value i got from them was the root cause (however, i wonder how they let the driver go into such a state - that's another matter).

Thanks,

Adrien

[tetedeiench]

I got 3 confirmation already the issue was fixed - expect a release in the upcoming hours !

[tetedeiench]

OCCT 6.2.2 released, with a workaround, so use this version please !