Since its original announcement at GTC 2014, the enterprise side of GRID has moved from a hardware-oriented product designed to sell integrated systems into a software product designed to be licensed for use on machines from approved vendors. The vendor list has grown considerably since that first announcement, as has the list of supported applications. Nvidia has stated a commitment to having two major feature update releases to GRID each year and is today announcing its August 2016 release.
At this point, the GRID-enabled GPUs are purchased for a qualified server like any other expansion component. Vendors are also much more accommodating to the idea of having GPUs in the servers, including having multi-GPU servers for large-scale vGPU deployments.
The main new and/or improved features in this release primarily deal with the ability to monitor and manage the individual vGPUs on a machine. Previously, overall usage could be monitored on a per-machine basis, but not effectively. You can use standard monitoring tools such as PerfMon, NVSMI and even GPUz, in addition to more specialized products like SysTrack from Lakeside Software. Monitoring and logging this kind of data, especially GPU utilization, will give admins a better idea as to who might need a slightly different configuration for their virtual machine. The hope with the improved management tools is that monitoring of resources can improve the user experience by making sure users have provisioned enough resources in their virtual machines to perform their tasks, and at the same time not be over-provisioned, thus wasting capital on excess machine power.
The August 2016 GRID update is available immediately. The new Tesla M10, a GPU better optimized for user density that was announced in May, is also available now.