nVidia GRID 2.0 – New GPUs and New Licensing Model

In September 2015, GRID 2.0 was announced during VMWorld.  The promise of current generation Maxwell-based GPUs, when combined with the potential to develop solutions for form factors that were incapable of housing a full-sized adapter, made for a very compelling proposition.  Creative architecture by Tier 1 vendors has provided a means to achieve a level of density that can change the scale of economies as it relates to hardware for graphically accelerated virtual solutions.  This does come with a considerable catch as additional licensing that wasn’t present in the GRID 1.0 model has come into existence.  Costs scale with the need for dedicated resources from the visualization perspective, and the end result may not be advantageous depending on the profiles required to support remote resources.  The breakdown of what each license targets from an application or use case perspective is clearly defined on nVidia’s site.  Care must be taken when balancing the available compute resources and the GPU solutions within a given platform.  In a traditional blade form factor that accommodates a single MXM form factor GPU, excess resources may be available to provision additional virtual desktop instances that do not leverage a GPU.  Conversely, rack mount form factors that may house multiple add-in cards have the potential to enable a visualization-heavy platform that does not offer enough compute to provide a positive and expected outcome.  It will be extremely interesting to see what further advancements are in store during the next two GPU architecture refreshes.

In the better late than never department, AMD finally offered their salvo in this space.  While competition can keep costs in check, there are some immediate concerns that stand out with respect to which competitive GPU AMD chose to bring to the contest.  The Tonga-based solution contained in the FirePro S7150 (standard and x2 models) is based on the Graphics Core Next 1.2 architecture.  It’s certainly a stable product that has been around for years, but its age is detrimental in the aspects of power consumption and performance against nVidia’s current-generation Maxwell solutions.  The peak density per GPU is equal for both offerings, as 512MB of VRAM is allocated per each session in the most basic virtual profile.   However, the performance edge will remain with nVidia for the foreseeable future.