To launch this blog, I will talk about my saturday at La Cantine with a GPU barcamp.
I have seen many interesting and passionate people thanks to the planners : François Lagunas and Eric Mahé.
I have seen some of the presentation after a small intro.
Comprehensive Sensing first, which is a fascinating new way to capturate our world (from classic picture to spectral). More information on nuit-blanche.blogspot.com.
CUDA, the "general" programming language for Nvidia GPU (waiting for OpenCL to have a common language for all hardware). It seems powerfull. However the language is a bit hard due to specifical programming of architecture, optimizations and parallelism.
Like you see in diagram below, the architecture of CUDA is based on multiple grid running one kernel (program). Grids contain blocs which contain thread. One thread apply kernel on one data (for example a float).
Warning : there is only one if else prediction by kernel.
A very interesting note is that code run in a determined time ! We can so create a real time system !
Another note is that approximation disappear since Nvidia G200.
And a presention on a work thesis about real time representation of "Positron emission tomography" (medical imagery) with specific problem of memory optimization (but finally common in programming of parallel architecture.
Waiting to have the presentation online.
Special link for www.geeks3d.com present this day.
I have seen many interesting and passionate people thanks to the planners : François Lagunas and Eric Mahé.
I have seen some of the presentation after a small intro.
Comprehensive Sensing first, which is a fascinating new way to capturate our world (from classic picture to spectral). More information on nuit-blanche.blogspot.com.
CUDA, the "general" programming language for Nvidia GPU (waiting for OpenCL to have a common language for all hardware). It seems powerfull. However the language is a bit hard due to specifical programming of architecture, optimizations and parallelism.
Like you see in diagram below, the architecture of CUDA is based on multiple grid running one kernel (program). Grids contain blocs which contain thread. One thread apply kernel on one data (for example a float).
Warning : there is only one if else prediction by kernel.
A very interesting note is that code run in a determined time ! We can so create a real time system !
Another note is that approximation disappear since Nvidia G200.
And a presention on a work thesis about real time representation of "Positron emission tomography" (medical imagery) with specific problem of memory optimization (but finally common in programming of parallel architecture.
Waiting to have the presentation online.
Special link for www.geeks3d.com present this day.
Commentaires