Intel wants to share the memory between the CPU and GPU

One of the largest stumbling blocks to generate a significant increase in the performances of the GPGPU, and more generally with calculations shared between the main frame and the graphics processor is related to the fact that it is necessary to transfer the data from the main memory of the machine to that of the graphics card where they will be treated and then the data is transferred in the opposite direction so that it can be used.
Intel, which continues to work on its Larrabee project decided to solve this problem with an innovative technique.
They will create a virtual memory shared between the two elements and in which each unit will be able to pick according to its needs without having to pass through the PCI-Express bus, a true bottle neck.
We do not know yet the details of the procedure, but it seems very promising on paper, we are waiting to see what Larrabee will give us, however it continues to be delayed.
As a reminder, Intel wants to go an opposite direction to that of ATI and especially to AMD by creating graphics cards which would be made up many cores of X86 processors put in parallel. In exchange to their loss of performances related to their lack of specialization, they would be programmable and usable as X86 processors and this would be much simpler to exploit for the developers.
The main issue here is related to the drivers who must ensure the performances suitable for the display by distributing the data with a maximum effectiveness in all the cores and without notable waste of time.
However, one knows from experience, Intel is rather bad at making graphic drivers.