Dedicated compute (OpenCL/CUDA/PhysX) card?

[BD2003]

So I picked up a GTX 750 for the hell of it, since it has good mining performance. No issues running the second GPU in the background mining away while I use the PC and the main GPU (GTX 770). I also have it marked as a dedicated PhysX card in the few games that use it.

Beyond that, is there any way to offload any other compute functions onto it? For stuff like tressFX, or anything else devs might come up with going forward? SLI seems like an outdated concept, since more and more of the GPU time will be dedicated to non-rendering tasks that don't need to be on the main GPU, and SLI is wasteful with costly memory and introduces all sorts of issues. There's also the issue of half of the die space on modern CPUs going to waste on an integrated GPU that just sits idle.

Perhaps this is something DX12 will address? But until then, anything I can do now to put the second GPU, or even the iGPU to good use?

 

[Makaveli]

I thought TressFX is amd only?

 

[BD2003]

I think it was at first, but hasnt been for a long time.

 

[bystander36]

TressFX works on both brands. It just uses standard DirectCompute, which both brands have support for. AMD just has better performance with DirectCompute.

As far as a dedicated DirectCompute card goes, I have never heard of such a thing.

I believe your best bet is just to SLI. It is not nearly the issue you seem to have come up with in your mind. While there are games which don't support SLI, or support it well, it is quite rare.

 

[BD2003]

TressFX works on both brands. It just uses standard DirectCompute, which both brands have support for. AMD just has better performance with DirectCompute.

As far as a dedicated DirectCompute card goes, I have never heard of such a thing.

I believe your best bet is just to SLI. It is not nearly the issue you seem to have come up with in your mind. While there are games which don't support SLI, or support it well, it is quite rare.

Well, it would be silly SLI-ing a 750 with a 770, no? I thought the cards needed to be matched.

 

[bystander36]

You can't SLI those cards. I never said you could.

 

[BD2003]

Well, I guess I'll reconsider SLI when the high end maxwell cards hit. Either way, as long as we've all got iGPUs in our CPUs, it'd be nice to be able to tap that power for something useful.

 

[Anarchist420]

TressFX works on both brands. It just uses standard DirectCompute, which both brands have support for. AMD just has better performance with DirectCompute.

titan actually has the best performance with direct compute although i guess AMD has better performance on average since AMD doesnt willfully damage most of their dies. nvidia makes me mad for forcing optimizations when they arent the only ones playing the games, for taking features away, and for screwing gamers over by telling them bs like there is a dichotomy between "gaming features" and "compute features".

intel, nv, and AMD feed their customers little better than crap when they could be selling the very best. they should have their IP taken away to never be granted IP again. hopefully, other people believe that too

 

[Makaveli]

intel, nv, and AMD feed their customers little better than crap when they could be selling the very best. they should have their IP taken away to never be granted IP again. hopefully, other people believe that too

O they sell you the best stuff just not in the consumer space its all in the enterprise space.

You gotta pay to play as they say! You want the best it will cost you!

If you expect them to give you the best possible cpu's they can make in the $200-$400 price range you are dreaming.

You do realize the spends 100s of millions on R&D, node improvments your not dealing with walmart here!

 

[sxr7171]

O they sell you the best stuff just not in the consumer space its all in the enterprise space.

You gotta pay to play as they say! You want the best it will cost you!

If you expect them to give you the best possible cpu's they can make in the $200-$400 price range you are dreaming.

You do realize the spends 100s of millions on R&D, node improvments your not dealing with walmart here!

That's funny because Walmart is bigger than Intel+AMD+Nvidia combined.

 

[Makaveli]

That's funny because Walmart is bigger than Intel+AMD+Nvidia combined.

Walmart having more money doesn't mean anything.

Coke also has more money than Intel+AMD+Nvidia!

 

[f1sherman]

titan actually has the best performance with direct compute although i guess AMD has better performance on average since AMD doesnt willfully damage most of their dies. nvidia makes me mad for forcing optimizations when they arent the only ones playing the games, for taking features away, and for screwing gamers over by telling them bs like there is a dichotomy between "gaming features" and "compute features".

intel, nv, and AMD feed their customers little better than crap when they could be selling the very best. they should have their IP taken away to never be granted IP again. hopefully, other people believe that too

I'll admit it's annoying, but how would you structure GPU business?

When there are people running business who are glad to pay thousands of $ for particular GPU feature crucial for their software (which btw often costs like new BMW),
and then there are those (so called consumers) willing to pay only few hundred dollars for that same GPU and mostly don't care about the feature?

 

[BD2003]

Is there literally no escaping the AMD vs Nvidia nonsense here? It has nothing to do with the topic of the thread.

 

[f1sherman]

We went little offtopic, but I see no AMD vs NV nonsense in this thread yet.
TBH it's you who went off topic 1st by mentioning SLI as it's in no way connected to your original question of offloading work to GTX 750.

In any case, beyond PhysX you can't offload general application or game workload to your secondary GPU ; you can only possibly run second instance of that same app.

Though there are some applications that can do exactly what you ask for, but I think they require Quadro/Tesla

http://us.acer.com/ac/en/US/content/nvidia-maximus-solutions

 

[BrightCandle]

Currently the best you can do is set it as the card for openCL and PhysX. I do envision a future where maybe things will be smarter and we can have multiple GPUs with different capabilities, but practically its not going to be soon.

 

[BD2003]

How does one set it as a card dedicated to OpenCL?

 

[Makaveli]

How does one set it as a card dedicated to OpenCL?

You select an application that uses Opencl

and in the drop down of which gpu to use you select the one you want?

Example Vegas Pro

 

[BD2003]

You select an application that uses Opencl

and in the drop down of which gpu to use you select the one you want?

Example Vegas Pro

Ah ok, so it's per app then, not a global setting like physx? Any games out there that let you select which card?

I'm honestly a little surprised GPGPU isnt pervasive by now...it's been years since CPU performance hit a wall, and so much time, energy and die space has gone into building compute capable GPUs into CPUs over the past few years. I'd think AMD especially would be at the forefront of this.

 

[Makaveli]

Ah ok, so it's per app then, not a global setting like physx? Any games out there that let you select which card?

I'm honestly a little surprised GPGPU isnt pervasive by now...it's been years since CPU performance hit a wall, and so much time, energy and die space has gone into building compute capable GPUs into CPUs over the past few years. I'd think AMD especially would be at the forefront of this.

We are now seeing a small amount of applications that use gpus now. However the hardware is always ahead of software, and I think it will still be sometime for it to mature.

 

[norseamd]

intel, nv, and AMD feed their customers little better than crap when they could be selling the very best. they should have their IP taken away to never be granted IP again. hopefully, other people believe that too

wrong that ip should be available for all

 

[norseamd]

was asking this exact same question in the cpu forum

how much gaming performance would a professional gpgpu add if the code was written for it