Posted by: Andy Trevor

The topic of Virtual PC and Workstations is a hot one in VDI at the moment.  For a number of years now we have had the capability of deploy vGPU technology using the NVIDIA K1 and K2 GRID boards.  Support was initially for Citrix, but VMware now have the same goodness that GRID can deliver.  The K1 card is the low end card designed to bring a small amount of vGPU love to virtual desktops where the K2 board is much more performant.  K2 is able to deliver some pretty cool results at the higher end of Virtual Workstation usage scenarios.  The performance of both cards is reflected in the price, with K2 being around double the cost of a K1. All of this goodness is provided license free on purchase of a K1 or K2 card. Where am I going with this you ask? This is all stuff we know. Bear with me while I move along.

The addition of vGPU to a VDI deployment adds a significant cost to what is already an expensive solution in terms of CAPEX.  Both Vmware and Citrix licensing costs are not insignificant, and the tin required to run such solutions is still not on the cheap side.  Again, all stuff we know, so why the rambling?

The K1 and K2 boards based on ageing Kepler technology are nearing EOL, in fact the end of 2016 will see the last of the from NVIDIA (check with your OEM as they may well have an earlier EOL date than this).  Fear not as towards the end of 2015 NVIDIA released GRID 2.0 based around the new Maxwell architecture, the new boards are the Tesla M60 and Telsa M6.  I will be concentrating more around the M60 as this is the form factor that most of the customers that we deal with here at Cutter will be looking at.  The M6 is an MXM form factor that I suspect will have little to do with.

The M60 is a beast of a board, the specs are very impressive as are the results, but that is not the point of this post. What is the point you ask? Well, the point is I think NVIDIA may well have shot themselves in the foot a little with the whole GRID 2.0 side of things.  I think they have done this on a number of fronts.

• Firstly, cost.  The M60 is reassuringly expensive to acquire.  One of my customers recently described it as eye watering.  They are 1.5 the cost of a K2 and near 3 times the cost of a K1.  This hike in cost does come with a hike in grunt I agree, but as you will see later down the page there is something missing.
• Secondly is the licensing model, which actually falls in under cost also.  It is no longer a case of buy a board and crack on with all it’s goodness. No, one now has to pay a licensing cost for the privilege.  Again, this post is not about the detail of this.  Costs of licenses can be found easily from NVIDIA or from us should you want a quote.  This is about the above increase in CAPEX and this increase in OPEX (albeit you can buy a perpetual license but you will need to buy support to keep up with the advancements)
• Thirdly and the main reason why I think they have shot themselves in the foot is around board density.  This has remained at 32 per board maximum at the lowest profile, sure those users get more grunt, but at a cost, a significant cost.

NVIDIA have completely forgotten about the entry level vGPU user.  The user that only needs a little bit of vGPU love to support basic Photoshop and the likes.  We see lots of this type of requirement in Education.  The cost of entry has risen 2.5 time in CAPEX and also now has associated OPEX costs.  These entry level users are the testbed that a lot of the oragnizations we work with use to judge the viability of GRID as a technology…..feeders for the bigger boys toys that GRID can handle.  This has been cut off at the knees by NVIDIA not providing entry level capability at an affordable price.  It has left the door open for the likes of AMD who have just hit the market with what on paper looks like a K1 and K2 type capability and cost boards.

This is all a little short sighted in my books.