OpenCL news

[BarsMonster]

It appeared that yesterday nVidia released OpenCL support for GPU, I am trying it now.
AMD promises that later this year.
Let's take a look at BarsWF performance with OpenCL Hope I'll need to write everything just once (Meanwhile latest AMD's StreamSDK didn't worked with Windows7 )

[Corni]

Do you know if there's a (substantial) speed difference between cuda and opencl for nvidia cards?

[BarsMonster]

Do you know if there's a (substantial) speed difference between cuda and opencl for nvidia cards?

I don't know, but downloading SDK and drivers right now, so I hope I will know today

[BarsMonster]

Love W7
I have both AMD & nVidia cards installed. Monitor connected to AMD 4870, I've extended desktop to my 8400GS although have no monitor attached.
Now I can have both AMD & nVidia working at the same time
No more hassle switching cards

[DarkPrince]

Hmm, I know that we have a real Windows 7 release CD at my school. Maybe I should ask for it so I can get win7

[IvanG]

Haven't tried nVidia's OpenCL by my own but have read that it's slower than CUDA. Not a surprise actually, compatibility costs some resources.

As for ATI's OpenCL I have serious doubts that it'll available any time soon. Also chances quite high than only OpenCL compatible GPUs will be HD5800s.

Shortly, I don't feel that OpenCL is ready to use -- for nVidia CUDA is already working perfectly while ATI going in very weird directions when coming to software.

But I'm curious to try OpenCL BarsWF .

[BarsMonster]

Here are the numbers:
CUDA: 26.73 MHash/sec
OpenCL: 25.38 MHash/sec (most of stuff is hardcoded, so actually it should be a bit slower)

[D3ad0ne]

So does this mean that programmers can start programming for OpenCL and it will be supported by anything that supports OpenCL? Both ATI(once they support it) and Nvidia without rewriting the code?

[BarsMonster]

So does this mean that programmers can start programming for OpenCL and it will be supported by anything that supports OpenCL? Both ATI(once they support it) and Nvidia without rewriting the code?

Yes. There still might be some minor differences between platforms(i.e. you still should think about optimal thread count for each platform e.t.c), and CPU implementation might be not fast enough (vs good SSE2 code).

Personally, I guess I am going to stick with CUDA for some 0.5 years to get mature OpenCL (better optimization, less issues, good CPU implementation), and then will work with OpenCL only.

[IvanG]

Here are the numbers:
CUDA: 26.73 MHash/sec
OpenCL: 25.38 MHash/sec (most of stuff is hardcoded, so actually it should be a bit slower)

Nice. I was more pessimistic about OpenCL .

Now only need to wait for ATI move.

[IvanG]

ATI finally released OpenCL for GPUs .

[BarsMonster]

ATI finally released OpenCL for GPUs .

So what this means for IGHASHGPU? Is there a chance, that hashcracking will utilize more than 78% of theoretic speed with ease?

OpenCL was never faster than native code... So, personally I doubt

[IvanG]

Yes, it's very unlikely that OpenCL code will be better than CAL/IL code. There was Brook+ built on CAL/IL layer, now Brook+ replaced by OpenCL but it's still built on CAL/IL layer.
It's like for CPUs -- we can write in asm (CAL/IL) or use C (OpenCL). If we're good in asm then asm code will always be better than C code. But writing in C is just easier than writing in asm.

So I'm planning to migrate to OpenCL (to avoid unnecessary work with CAL/IL and potentially to write only one kernel for both nVidia & ATI GPUs) but performance won't change in better way unless algorithm will change in better way.