NVIDIA has just announced at its GTC conference that the NVIDIA Kepler* graphics processor (GPU) has been designed to support virtualized environments which are the basis of cloud computing. GPUs are traditionally very hard to integrate in a virtual environment because GPUs have traditionally been designed around performance, which often resonates with a “direct to the metal” programming mentality.
These days, GPUs are fast enough to add a thin hardware layer that translates virtual addresses to physical addresses, along with virtualizing every graphics hardware commands.
With this new design, and tons of software around it, NVIDIA GPUs are now able to work seamlessly in a virtualized environment. GPUs can now be shared among many users, and they don’t even have to be connected to a physical video output anymore. GPU sharing also enables the maximum utilization of the GPU hardware because single users often don’t fully utilize all the computing power of those chips. This utilization rate makes the GPU more power-efficient than it otherwise would be.
At GTC, NVIDIA has demonstrated how its new GPU could be virtualized in a Citrix environment. If you’re not familiar with it, Citrix is a very popular virtual environment for the enterprise. In the demo, the real-time 3D graphics were running much smoother than they would normally do in such a virtual environment. In theory, every single GPU application should run within Citrix now. To demonstrate this capability, NVIDIA has invited Industrial Light and Magic (ILM) on-stage to demonstrate how a special effects supervisor can access terabytes of data from a remote location, on a small laptop.