cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

Concurrent kernel Execution using multiple command queues

I use Nvidia quadro FX880 card. Does the card support concurrent parallel execution if multiple command queues are created...? Got answers on google that it doesn't, but if it doesn't then it should also not support task parallelism which is executing two or more kernels concurrently...! Please clear my doubt about multiple command queues..!

0 Likes
7 Replies
nou
Exemplar

why do you ask on AMD forum about nVidia HW? go to ask nVidia.

himanshu_gautam
Grandmaster

As far my limited knowledge goes, NVIDIA supports concurrent kernel execution only in Kepler-architecture  cards. So go check what architecture your quadro card currently has.

Also, Whether NVIDIA supports this concurrent execution in OpenCL -- is totally a different question.

You can keep asking NVIDIA about this. They have not even released OpenCL 1.2 yet.

So,

If you want to stick to OpenCL -- you better be with AMD. Buy an AMD 7xxx series card and it will rock your code.

Good luck!

Hey Nou and Himanshu,

   I know  that I am asking about Nvidia cards on AMD forums. People on Nvidia forums don't answer questions related to OpenCL as we all know. I really regret on buying an Nvidia card when I knew they love only CUDA and not OpenCL.

Please try to help me out of this.

     As per what you say Himanshu, does it means if kernels are enqueued using clEnqueueTask on Nvidia cards then even they would execute in a sequential manner, when they are actually meant for task parallelism...? also does this mean that Nvidia cards support only Data parallelism and not task parallelism.

     And I guess Nvidia released OpenCL 1.2 yesterday... But then will that add concurrent kernel execution to the card..?

0 Likes

I think clEnqueueTask() will run your kernel with 1 work-item... Oops... Dont even try it on GPUs.

Check the OpenCL Spec on this. if it executes with only 1 work-item -- then forget it on GPUs.

This API is probably meant for embedded platforms (or) CPU based platforms - where it does not make sense to spawn large number of workitems (i.e. if the CPU RT is not smart enough to map the large number of workitems onto small number of processing cores)

GPUs are data-parallel.

And, Thanks for the news that NVIDIA released OpenCL 1.2. Finally, they have woken up to reality.

0 Likes

Hi Shreedhr,

Where did you find info on NVIDIA releasing OpenCL 1.2 support?

I dont see it in their page...(or) any press relases on it....

0 Likes

I actually read it on the facebook OpenCL group page, and I guess it's not released it's just announced. Sorry for the inconvenience...!

0 Likes

O...Thats nice...Facebook OpenCL page...Good to know this.

I dont see the announcement or press release anywhere else...May be, some frustrated developer is spinning out rumours. Unless we see the official announcement, we cant really judge.

Anyway, Let me go check their FB page to see what this is all about. Thanks!

0 Likes