GPU virtualization in XLcloud

Jun 20 2012

During the last GTC, NVidia has announced its new generation of cloud computing-enabled GPU chipset, code name KEPLER. Of particular interest with regard to project XLcloud is that for the first time an NVidia technology enables GPU virtualization in the cloud. This innovation addresses specifically the usage of the graphical capacity of GPUs in the cloud. Cloud gaming and more generally cloud graphics have been announced as strategic applications for NVidia and so reenforces our initial vision of providing deported rendering in the cloud.

Technical Overview

The cloud-enabled VGX technology developed by NVidia around its KEPLER chipset consists of two parts:

  • The KEPLER chipset itself integrates a low-latency H264 encoder. This encoder enables the GPU to handle directly the compression of the produced graphical video stream. The announced performance is the production of four streams of 30 frames per seconds at 720P quality.
  • A software hypervisor that enables remote control of the GPU card on the server from the client side.

KeplerArch.PNG 

As described in the above diagram, the hypervisor is able to establish a connection with several virtual machines in a cloud environment. Each virtual machine gets a dedicated stream resulting from the graphical compute of the GPU that is directly produced by the frame-buffer of the board.

It is unclear at this time what will be the level of openness of the software layer or the ability to use CUDA compute capacity of the GPU through this hypervisor as NVidia’s strategy has been not to disclose its GPU's internal architecture so far, hence forcing application developers to use NVidia's proprietary libraries as opposed to using open-source standards like CUDA or OpenGL libraries.

The product line of NVidia comes as a server-enabled board named NVidia GeForce Grid.

KeplerSpec.PNG 

How does it impact XLcloud ?

This technology aims to provide a neat solution for providing remote graphic rendering capabilities on a virtual machine. As a result, it may impact directly the work to be done in the XLcloud project. In particular, we will consider taking advantage of KEPLER depending on the level of openness and compliance with standards of the software artifacts. Far from striking XLcloud of irrelevancy, KEPLER reinforces our vision and provides additional opportunities to bring remote rendering in the cloud. In addition, we will strive to demonstrate use case scenarios, like scalable graphics applications, where a problem cannot be solved by a single instance of virtual machine but by a cluster of virtual machines. Another interesting domain will be to investigate we can provide a virtual cluster on demand capable of dispatching several small 3D remote viewing sessions per GPU for a very large number of concurrent users (3D VDI scenario).


This wiki is licensed under a Creative Commons 2.0 license
XWiki Enterprise 5.4.6 - Documentation - Legal Notice

Site maintained by