GTX1060 Linux drv v430 Error compiling kernel BAD_WORK_UNIT

[m1geo]

Hey, thanks for the confirmation. I've been doing some digging, and I notice something weird. As the GPU changes up through the power levels, the FAHBench https://fahbench.github.io/ benchmarker falls over as soon as the GPU enters power level 3 (P3). Now, I haven't ever overclocked the GPU (GTX 1070) and PSU is more than big enough to run the 150W dissipation.

The GPU will happily sit running a hacked about version of the matrix multiply (just the multiply operation in a infinite loop, recompiled):


Again, thanks for the help.

[Joe_H]

The GPU folding core does not use CUDA, it uses OpenCL.

[m1geo]

Yeah, I realise this. It was more a proof of concept that the GPU is there, will talk to the PC, and will run something without crashing.

[m1geo]

Some more debugging with FAHBench https://fahbench.github.io/...

The GPU benchmark runs to about 10% before falling over with either "NaN" error or some random exception (usually clEnqueueMapBuffer).





I'm not too sure what to do with this information, or how to debug further...
Thanks...

[toTOW]

It's not a sign of good shape of your GPU ...

[m1geo]

That's what I thought.

I'm new to GPUs. I'm an electronic engineer, so I don't know the details of GPUs, Nvidia settings, parameters, etc., but I have a pragmatic considered approach.

What I have learned this evening is that reducing the power limit down makes things behave, and I can complete the test.



My current working theory is that there is either a power issue or a clock speed issue which the lower power limit prevents the GPU from entering. The GPU came from a friend, but maybe he dabbled with overclocking it in the past and didn't remember. Thanks all.

[ipkh]

I've never managed to get a bad GPU to work via downclocking. And I have sent at least 4 cards back for warranty service in the past 5 years. 24/7 folding just has a habit of revealing faults with graphics cards.
You should definitely contact the manufacturer about warranty status on that card.

[m1geo]

Thanks for the heads up.

Using Uengine Heaven https://benchmark.unigine.com/heaven looping, I am able to get the GPU performance at
[*]Graphics: -70
[*]Memory: +500

However, a 10 minute FAHBench session doesn't like that. For FAHBench, I need to run:
[*]Graphics: -90
[*]Memory: +300

The card is an Asus ROG Strix GTX 1070 O8G with the factory heavy overclock, so I guess I'm just reducing that default a little. If I had a Windows key, I'd check to see how it performed on Windows. Maybe worth a shot.

Thanks for the help/advice.

[kevinjos]

m1geo - I do not see the same compile error reported by florinandrei. I think we may be dealing with separate issues. Do you consistently see the BAD_WORK_UNIT warning?

florinandrei - have you tried to compile the cuda samples as shown by m1geo above?

[toTOW]

Rule number one to overclocking with FAH : don't touch the VRAM clocks ! It creates more issues that it adds performances.

However factory overclocks should work ... it they don't, then the card need a RMA.

Does the card runs fine in Furmak with manufacturer default clocks ?

[m1geo]

I finally found my issue. It's bizarre! One of the fan bearings has failed. When the controller tried to spin the fans up, the fan would spin a bit, then jam, then drag the 12V rail down on the GPU. That caused all kinds of weirdness. Simply unplugging the one fan and the card works fine. I have ordered 3 new fans. Thanks for the patience and the advice!

[Neil-B]

Wow … damn good spot/catch … at least fans are (I believe) cheaper than new GPU card !!

[toTOW]

That was a nasty one ...

[kevinjos]

Right on, good catch! I'm curios if florinandrei was able to sort out their compile error. Are there a set of standard programs to test the system's ability to compile OpenCL code?

[Roadpower]

Nice catch indeed.