What is CPU/GPU memory sharing?

Early this morning, the NVIDIA official theme meeting, the old Huang announced the next generation of GPU, code-named Pascal, but also will join Nvidia up to the latest Nvlink memory sharing technology. Over the years, the traditional CPU, GPU can not share video memory, physical memory is the first time the old yellow break.

So how does this work? According to the Nvidia official, the actual use requires Nvidia Nvlink memory technology, 3D virtual memory two technologies.

Simply put, the existing unified storage allows graphics graphics cards to access and use GPU memory (video memory). For example, a software developer can specify the resource allocation between the two in the program code or define an application usage.

However, the unified storage limit CPU, the GPU can only access the use of their own memory for data calculation and exchange, and therefore need to remove the limitations between the two. For example, the APU chip produced by AMD.

For example, a PC computer, configured with 32G DDR3 memory, and equipped with one or two graphics card, each video card configuration 6G or 8G GDDR5 memory.

Therefore, the computer accumulated maximum memory capacity of 48G, of which 16G from graphics graphics card memory, for the CPU, GPU, both can read access to the highest 64G DDR3/DDR4 memory capacity.

The NVIDIA Nvlink memory technology can provide 5 to 12 times times the speed of the PCI Express interface's task data sharing, but is mainly used for some high performance computer equipment.