Drivers for linux are rubbish

[Jimbo]

Probably on future generations nvidia will be recovered of several hardware and mainly manufacturing errors they are suffering. But certainly the fermi generation is a complete disaster.

Anyway they remain *very* strong on some points:

- Cuda 3.0 is better than openCL
- VDPAU

[Hans]

Probably on future generations nvidia will be recovered of several hardware and mainly manufacturing errors they are suffering. But certainly the fermi generation is a complete disaster.

Anyway they remain *very* strong on some points:

- Cuda 3.0 is better than openCL
- VDPAU

In what areas is Cuda better than OpenCL? I have just finished a project with OpenCL from AMD's stream SDK 2.1. Last year I did a project with CUDA 2.1.

[pingufunkybeat]

I still don't understand why we are discussing whether the forum posts on the internet somewhere are written by nvidia users or amd users. It's a useless discussion.

Go to your local retailer, get a graphics card, test it on your computer, and if you don't like it, give it back.

It will take you half a day, instead of making theories on forum posts for weeks.

[Jimbo]

In what areas is Cuda better than OpenCL? I have just finished a project with OpenCL from AMD's stream SDK 2.1. Last year I did a project with CUDA 2.1.

I have read 2 papers, only by curiosity and not in detail, i don't program with cuda or opencl. I didn't save them, but it will be easy to find them on google again.

Mainly cuda 3.0 offers more facility, more integrated functions, and more performance. Some projects already implemented on cuda 3.0 were ported to openCL with worst results and an extra difficulty while implementing/optimizing them.

[Panix]

I still don't understand why we are discussing whether the forum posts on the internet somewhere are written by nvidia users or amd users. It's a useless discussion.

Go to your local retailer, get a graphics card, test it on your computer, and if you don't like it, give it back.

It will take you half a day, instead of making theories on forum posts for weeks.

What makes you think you can buy a card and bring it back because 'you don't like it?' Someone already posted this, why are you telling ppl to do this, still?

There's some big box stores where you might be able to do this but the cards have a huge markup compared to the typical price of an online vendor for the same or similar card.

Anyway, I'm resigned to give up criticizing. Most of my future posts will be reserved for specific questions.

[Hans]

I have read 2 papers, only by curiosity and not in detail, i don't program with cuda or opencl. I didn't save them, but it will be easy to find them on google again.

Mainly cuda 3.0 offers more facility, more integrated functions, and more performance. Some projects already implemented on cuda 3.0 were ported to openCL with worst results and an extra difficulty while implementing/optimizing them.

Okay, thx for your answer. I might look into it.
From my experience, OpenCL's and CUDA's kernel language is very much a like. Its just that OpenCL is harder to set up before the exectution of the kernels. You have to set up some vendor specific parameters like kernel size, which is totally different between nvidia and ati hardware.

Though you are actually able to tweak your OpenCL app to vendor specific optimizations before the kernel executions. But that part is harder, than just use CUDA with nvidia hardware. I still prefer OpenCL though, because AMD hardware tends to be faster than nvidia in OpenCL due to a higher number of mikro-processors within the GPU.

If you are only working with nvidia hardware, CUDA might be better. But if you compare CUDA on nvidia hardware vs OpenCL on Ati hardware, you might see better performance with OpenCL due to stronger hardware. At least thats my experience.

[pingufunkybeat]

There's some big box stores where you might be able to do this but the cards have a huge markup compared to the typical price of an online vendor for the same or similar card. questions.

I personally could live with a $15 markup if this saved me months of arguing on the internet, and STILL not resolving the only relevant question.

[mirv]

Buying a card to test is not viable for everybody. Not everybody has access to some kind of big box store, and even then I don't know any store that would take something back simply because "you don't like it".

[Jimbo]

Okay, thx for your answer. I might look into it.
From my experience, OpenCL's and CUDA's kernel language is very much a like. Its just that OpenCL is harder to set up before the exectution of the kernels. You have to set up some vendor specific parameters like kernel size, which is totally different between nvidia and ati hardware.

Though you are actually able to tweak your OpenCL app to vendor specific optimizations before the kernel executions. But that part is harder, than just use CUDA with nvidia hardware. I still prefer OpenCL though, because AMD hardware tends to be faster than nvidia in OpenCL due to a higher number of mikro-processors within the GPU.

If you are only working with nvidia hardware, CUDA might be better. But if you compare CUDA on nvidia hardware vs OpenCL on Ati hardware, you might see better performance with OpenCL due to stronger hardware. At least thats my experience.

The papers i looked were from universities, some communications and math projects. That plp are not ati or nvidia fans XD but they marked that cuda 3.0 was better. Obviously openCL is the future because is open, and continously improved. But nowadays in the performance / easy implemaentation cuda 3.0 seems to win. If i don't remember bad, openCL has some memory management problems which makes worse the performance, and to partially avoid these problems you have to program very hard!.

[Panix]

I personally could live with a $15 markup if this saved me months of arguing on the internet, and STILL not resolving the only relevant question.

It's way more than $15.

One can buy a card and not be happy with it. They can send it back and get a re-stocking fee. That's if they can get away with saying 'they don't like it.' Or they can pay from $50 and up for the same card at a big box store.

Anyway, if there is no return option for 'not liking the card,' they can sell it but will take a hit for that as well. It's not so simple as you're implying.

But, yes, one should have the card and try it for the experience as possible. Easier to discuss it, too, I gather...

Okay, I'm done ...leaving the discussion unless I have a technical question.