Virtual desktop infrastructures and virtualized apps are becoming common in today's businesses as organizations move away from desktop PCs and deliver applications or entire virtualized PCs over the network. Even basic productivity apps like Microsoft Office and the Windows 10 operating system now expect GPU acceleration, though, and Nvidia thinks virtualizing today's software without a GPU accelerator running underneath is a fast track to a sub-optimal user experience. To solve this need for GPU acceleration for everyone, Nvidia is offering system administrators the ability to put up to 64 virtual users on a single shared graphics card with the new Tesla M10.
The Tesla M10 puts four "mid-level Maxwell" GPUs on a single card, each paired with 8GB of GDDR5 RAM for a total of 32GB. Going by the stream processor count, it seems possible those are GM107 GPUs, as found in the GeForce GTX 750 Ti. Nvidia says that putting two Tesla M10s in a server lets that node meet the magical 100-user-per-server metric it claims system administrators are looking for when planning virtual infrastructure deployments. As the diagram above suggests, the M10 adds high-density GPU resources to Nvidia's existing Tesla product line: the blade-server-ready Tesla M6 and the high-performance Tesla M60.
The M10 is a new brick in the foundation of the Grid virtualization stack. For administrators to actually use the M10, they'll also need to purchase per-user licenses for their deployments. Nvidia already simplified the types of licenses it offers for Grid earlier this year. To give organizations more flexibility in their licensing requirements, the company is now offering annual licensing that's similar to the subscription model Adobe uses for its Creative Cloud apps. Under this model, an annual virtual app license will cost $10 per concurrent user, a virtual PC license will cost $50 per concurrent user, and a virtual workstation license will cost $250. For more information, check out Nvidia's Grid portal.