GPUPI - SuperPI on the GPU

[_mat_]

Yes, it's cleaner code with improved comparability between different devices and the OpenCL path is now implemented correctly as it always should have been.

The different OpenCL drivers produce closer results (although AMD OpenCL 1.2 = AMD APP SDK 2.9-1 is now the best choice in all categories), Batch Size and Reduction Size are not as picky as it was before and on NVIDIA cards the OpenCL implementation comes very close to the CUDA implementation, which indicates that everything is done right now.

The bottom line is GPUPI is now much better as a benchmark in general. I would have done this with GPUPI 1 already, if I could have. But I wasn't good enough at OpenCL coding and mathematics back then (OpenCL is a brutal beast though).

The good news is that something like this won't happen again. I will not touch the algorithm anymore, because it's pretty much maxed out the way I do it. The next step is an OpenMP path that gets rid of the OpenCL implementation, but that's many months away and will not overrule current results. The CPU path will be split into OpenCL and OpenMP (or Native, don't know yet), so no rebenching necessary. The new path will make use of AVX and whatever comes next to support the hell out of everything that comes in my way.

I know the XTU coders and they don't seem to be interested in overclocking, let alone competitive oc. They just do their job and as far as my experience with XTU SDK goes, it's not entirely a good one (sry guys). I really try to do things differently with GPUPI. I want it to be on the bleeding edge too, but I wouldn't have introduced the speedup with 3.3 if everybody in this thread would have stood against it. As I already said: GPUPI should first and foremost be fun to bench.

[Aleslammer]

Said my say!

Was going to see how the new one tested on CPU got this.

 

 

[Splave]

ran on Titan V this morning. Old version 3.5**s now 2.78*s on 1B and 4m30s to 2m59s on 32b.  saving and upload working fine on 3.3.0.0 for it.

 

thanks for the update. maybe 32b is turning into a more viable option for global points now that its much faster?

[Alan_Alberino]

110pts with 5200MHz 7700K, seems it's good

[_mat_]

9 hours ago, Splave said:

maybe 32b is turning into a more viable option for global points now that its much faster?

Would definitely be one of the hardest GPU categories to give points, right? GTX 1080 Ti at 3000 MHz for over 4 minutes anyone?  

Edited April 6, 2018 by _mat_

[_mat_]

9 hours ago, Aleslammer said:

Was going to see how the new one tested on CPU got this.

It just stops when compiling the kernels. Well, that's bad. Please try this beta version of 3.3.3 that adds a lot of detailled debug log messages to the initialization of CUDA and OpenCL:

https://www.overclockers.at/downloads/projects/GPUPI 3.3.3 Beta.zip

Please post the contents of the debug log, thank you!

[unityofsaints]

Is it possible to fix RAM detection? At the moment GPUpi detects effective speed but HWBot subs are in mhz (half the effective speed). Thanks for all the work you put into this @_mat_

[Splave]

nevermind im dumb

Edited April 7, 2018 by Splave

[Leeghoofd]

Sorry Matt but this new version is ridicilous in performance... 5200 beating 6900MHz, nothing to do with tweaking OS, finding right CL driver...  You provided the boost for them

 

If it was me this version could not be used on anyhting older than e.g. Kaby/Skylake X or 10 series nVIDIA GPUs

 

Those that benefit won't complain, those which LN2 scores get beaten by  air setups have a valid point for removal...

 

Like Allen said maybe we need to move up a step and go for the heavier calculations.

Edited April 7, 2018 by Leeghoofd

[Leeghoofd]

10 hours ago, Alan_Alberino said:

110pts with 5200MHz 7700K, seems it's good

What does it prove in the sense of your overclocking capabilities?

 

I would urge everybody to stop benching the new version till the HWBOT team sorts out a viable solution for the usage of the new optimized version!!

 

[Splave]

I remeber not but a year or so ago I was required by the mods to prove my out of line efficiency in r15.... I was using a newly released version that scored better. Same thing happens with xtu. 

The people that this caters to is the people that are actively benching. Imo that's the people that matter. 

Just my opinions 

[Mr.Scott]

1 hour ago, Splave said:

I remeber not but a year or so ago I was required by the mods to prove my out of line efficiency in r15.... I was using a newly released version that scored better. Same thing happens with xtu. 

I have to agree.

GPUPI is no different.

Sorry, but that's the way it goes.

Consistency is still going to be a problem at HWB?

Edited April 7, 2018 by Mr.Scott

[Aleslammer]

15 hours ago, _mat_ said:

It just stops when compiling the kernels. Well, that's bad. Please try this beta version of 3.3.3 that adds a lot of detailled debug log messages to the initialization of CUDA and OpenCL:

https://www.overclockers.at/downloads/projects/GPUPI 3.3.3 Beta.zip

Please post the contents of the debug log, thank you!

Thanks, attached file.

HWiNFO64.DBG

[havli]

12 hours ago, Leeghoofd said:

Sorry Matt but this new version is ridicilous in performance... 5200 beating 6900MHz, nothing to do with tweaking OS, finding right CL driver...  You provided the boost for them

 

It is simple, just rebench the 6.9GHz CPU and done.

Those who want to stay on top must rebench stuff regularly anyway. For example in 3D with every new 115x platform launch rebench of all 3dmarks up to 06... and new HEDT = rebench of Vantage and later.

[Leeghoofd]

You assume everybody still has their gem CPUs. It is not only about latest generation processors,

Anyway your reference to the 3Dmarks is not valid as you just add more CPU horsepower to the benchmark. The 3Dmark06 or Vantage algorythm doesn't change here. Which is the case here, GPU PI has drastically been altered and is not the same benchmark it was before. If  Futuremark releases a version which boosts a 30% increase I wonder if you would just enjoy people beating your LN2 scores running stock GPUs.

 

At the moment removing all points scores for 2D one by one... and deleting. We will set up with Matt a new benchmark for this. But these scores will not be compared with the old ones, all 3.3s will be deleted!!

Edited April 7, 2018 by Leeghoofd

[shar00750]

Is this includes gpupi for gpu or just for cpu?

[Leeghoofd]

Both, I quickly tested 3.2 versus the new one, sorry this is a no go

[Mr.Scott]

2 GPUPI benches is ridiculous.

Nobody will ever bench the old one again with the newer, easier one out.

It will be a dead bench.

[Alan_Alberino]

12 minutes ago, Mr.Scott said:

2 GPUPI benches is ridiculous.

Nobody will ever bench the old one again with the newer, easier one out.

It will be a dead bench.

This... Old versions are dead since 3.3 was launched... Except you make 3.2 mandatory (Which will leave new HW users without support), everybody will prefer to bench 3.3

[bigblock990]

28 minutes ago, Leeghoofd said:

Both, I quickly tested 3.2 versus the new one, sorry this is a no go

Can you please explain your plan then? If 3.2 and older is one category and 3.3 is a new category will both earn global points or just one?

I guess if you must seperate them out, then have 3.3 with globals and hardware, then 3.2 older only hardware points. Everyone will bench 3.3+ from now on anyways.

Edited April 7, 2018 by bigblock990

[Leeghoofd]

I already talked to Matt today and this version will probably be called GPUPi4 and we will have to decide on other benchmarks for it. 

 

Let it be clear that I will not let this benchmark version interfere with scores of previous versions, not on my watch. Whatever happened in the past does not interest me at al.

After some testing and debating I decided that these boosted algorythm scores will not be included in any current GPUPI ranking. Hate me for it, I don't care. The created stir just proves again that people just prefer easy points and not a challenge.

 

@bigblock990You can bench whatever version you like, but till now only 3.2 or below will give you globals and hardware points. 3.3 will only give you pleasure and nothing in return at the Bot, untill a new benchmark is introduced.

 

 

[Leeghoofd]

48 minutes ago, shar00750 said:

Is this includes gpupi for gpu or just for cpu?

Both will not be accepted, sorry

[Leeghoofd]

38 minutes ago, Mr.Scott said:

2 GPUPI benches is ridiculous.

Nobody will ever bench the old one again with the newer, easier one out.

It will be a dead bench.

At least people will still benefit from their hard work done in the past, which is completely eradicated by this new version...  Stuff has been handled wrongly in the past, this time it might be my time to make the mistake, yet carry full repsonsability.

Version 3.2 will still be used for 100M and 1B for CPU / 32B for GPU. Nothing more and nothing less!

 

Edited April 7, 2018 by Leeghoofd

[MetalRacer]

Why not do a poll and let the community decide?

[Splave]

1 hour ago, Leeghoofd said:

I already talked to Matt today and this version will probably be called GPUPi4 and we will have to decide on other benchmarks for it. 

 

Let it be clear that I will not let this benchmark version interfere with scores of previous versions, not on my watch. Whatever happened in the past does not interest me at al.

After some testing and debating I decided that these boosted algorythm scores will not be included in any current GPUPI ranking. Hate me for it, I don't care. The created stir just proves again that people just prefer easy points and not a challenge.

 

@bigblock990You can bench whatever version you like, but till now only 3.2 or below will give you globals and hardware points. 3.3 will only give you pleasure and nothing in return at the Bot, untill a new benchmark is introduced.

 

 

This sounds ok. But when will the new one get points rev 8? I may be in the grave when that is released. 

In regards to mat i hope he doesnt take this as disrespect by hwbot. He has made probably the most secure benchmark on hwbot and continues to improve it for no personal gain. Hope to see the new version get treated well and not become black hole bench 2.0 bottom of the list to die.