- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi everyone,
I am evaluating the possibility to use the TESLA modules for parallel processing as a relative cheap solution compared to using Open MP and similar.
I went to the website of Nvidia and I found that there is a new FORTRAN called CUDA FORTRAN from the Portlang Group, which supports the TESLA MODULES programming. Isthere any possibility to use the Intel FORTRAN Compiler on such a technology? Maybe a library?
I would appreciate a lot your feedback
I am evaluating the possibility to use the TESLA modules for parallel processing as a relative cheap solution compared to using Open MP and similar.
I went to the website of Nvidia and I found that there is a new FORTRAN called CUDA FORTRAN from the Portlang Group, which supports the TESLA MODULES programming. Isthere any possibility to use the Intel FORTRAN Compiler on such a technology? Maybe a library?
I would appreciate a lot your feedback
Link Copied
4 Replies
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Not that I am aware of. The PGI "CUDA Fortran" translates the Fortran code to CUDA C and then runs it through that processor. They sell it separately from their regular Fortran.
The approach Intel is taking is to extend Intel Fortran (and C++) to allow offloading of code to an Intel "Many Integrated Core" device. You can read more about it here.
The approach Intel is taking is to extend Intel Fortran (and C++) to allow offloading of code to an Intel "Many Integrated Core" device. You can read more about it here.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi
Mostly irrelevant to the question asked here but one can use libraries such as CUDA-Blas which don't require CUDA-Fortran. How much useful that is and/or its simultaneous usage with OPENMP etc. etc., I can't comment.
Abhi
Mostly irrelevant to the question asked here but one can use libraries such as CUDA-Blas which don't require CUDA-Fortran. How much useful that is and/or its simultaneous usage with OPENMP etc. etc., I can't comment.
Abhi
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi I don't want to use simultaneously use the OpenMP and GPU technology even because they are not comparable. Currently OpenMP needs a very high bandwidth link between each computer/node in the cluster. Using a GPU this high speed/high bandwidth is embedded on the motherboard, since it uses the same channel that communicates with the PCIExpress ports. PNY is using the QUADRO videocard to provide a high computing package, consisting of 4 NVIDIA QUADRO with about 448 cores and a power of about 512 Gflops in double precision.
In order to attain the same performances on an MPI cluster you need a smartware connections (very expensive) and several cpu/cores. The bottleneck in this case is the way the different computers are working through the MPI.
Unfortunately the HPC machines are still very expensive and the solution proposed by NVIDIA looks like to be feasible on a financial point of view.
Just as an example to have a decent Superdome II machine, you need about 30 times more the investment of a TESLA machine. Of course the performances are not comparable, but the gap is not too much distant each other.
Anyway the suggestion to use CUDA-Blas libraries makes sense. I didn't know about their existence. Thanks a lot for your comment Abhi.
Kindest regards,
Gianfranco
In order to attain the same performances on an MPI cluster you need a smartware connections (very expensive) and several cpu/cores. The bottleneck in this case is the way the different computers are working through the MPI.
Unfortunately the HPC machines are still very expensive and the solution proposed by NVIDIA looks like to be feasible on a financial point of view.
Just as an example to have a decent Superdome II machine, you need about 30 times more the investment of a TESLA machine. Of course the performances are not comparable, but the gap is not too much distant each other.
Anyway the suggestion to use CUDA-Blas libraries makes sense. I didn't know about their existence. Thanks a lot for your comment Abhi.
Kindest regards,
Gianfranco
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
ThanksSteve for the link.
I read the info and I found them quite interesting. I don't know the state of art of the MCI hardware, but on some forums the main perception is that Intel just started in this new field, but it is still far from being competitive with NVIDIA.
The fact that the scalability of the code is ore easy on MCI machines is quite appealing. I hope that the costs will be appealing as well.
Running my codes now at University is costing too much to be in business. I needa in-house solution. Let's leave and see for now.
Thanks again for the infos.
Kindest regards,
Gianfranco
I read the info and I found them quite interesting. I don't know the state of art of the MCI hardware, but on some forums the main perception is that Intel just started in this new field, but it is still far from being competitive with NVIDIA.
The fact that the scalability of the code is ore easy on MCI machines is quite appealing. I hope that the costs will be appealing as well.
Running my codes now at University is costing too much to be in business. I needa in-house solution. Let's leave and see for now.
Thanks again for the infos.
Kindest regards,
Gianfranco
Reply
Topic Options
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page