nobeastsofierce - Fotolia

Virtual GPU is nearly ready for the enterprise masses

GPU virtualization is traditionally only for virtual desktop users that need to run resource-intensive applications, but more competition among vendors could change that paradigm.

Get ready for graphics processing units to become much more common across the enterprise.

One of the announcements from Citrix Synergy that didn't get a lot of coverage is that XenServer 7 became the first hypervisor to support ultra-high-definition, 4,000-pixel virtualized computing with Intel Iris Pro Graphics cards. Citrix's competitive advantage there is surely short-lived, as you can expect to see VMware add support for Intel Iris Pro Graphics around the time of VMworld 2016 in late August.

This news is the latest evidence of the overwhelming trend toward more widely-available and affordable graphics processing units (GPUs) that IT can slice up into virtual GPU (vGPU) for its high workload users.

The prices were too darn high

A few years ago, NVIDIA's GRID vGPU was the only platform around that could support the emerging use cases for 3D graphics virtualization. NVIDIA capitalized on this by attracting large customers willing to pay large sums of money to support high-end VDI use cases. But the relatively high cost per slice of a virtual GPU meant that the low to midrange users were left behind.

Finally, there is competition over midmarket GPU virtualization.

As time passed, though, IT realized that more applications, and even the Windows operating system itself, used GPU at an increasing pace. Apps such as Office and Internet Explorer now require some level of GPU to look and feel normal during a virtual desktop instance. Now, users without any graphics acceleration are saying the same thing IT has always heard about how their virtual desktop feels different than their PC.

NVIDIA tried to address that problem -- and increase its revenue -- by adding in software licensing tiers for its GRID GPU cards. High-end user licenses cost more, which grants them more resources. Companies pay less to support knowledge and task workers, while still providing them with enough vGPU to make their virtual desktops perform a little nicer. The problem is that GRID GPU cards are still too expensive for most organizations to use for the entire workforce. You can only slice the GPUs up so much, and a cheap software license means nothing if the hardware costs are still prohibitively high.

Don't get me wrong, because I've never felt that the enterprise should not support graphics virtualization. Every VDI deployment should incorporate it to some extent, and many do. But when your IT budget comes in higher than originally thought, guess what goes first? It isn't storage -- IT needs to have that. It isn't networking or thin clients, either. The first things you cut from the budget are what you don't absolutely require, and that's why graphics virtualization often gets the axe.

More affordable virtual GPU

Today, things are changing quickly. AMD put pressure on NVIDIA by releasing a product named Multiuser GPU, or MxGPU, that can support 64 users for $8,000. That's just over half the cost of NVIDIA's most comparable product. AMD MxGPU cards are available in three models through original equipment manufacturers Hewlett Packard Enterprise (HPE), Dell and SuperMicro. The basic MxGPU model is the S7150, which retails at $2,399 and can support 16 users per card. The S7150 x2 model includes double the resources, is slightly more power efficient and can support 32 users per $3,999 card.

NVIDIA responded by releasing a lower-end card, the Tesla M10, which significantly narrowed the price gap between the two competitors. The M10 card includes four GPUs that IT can split among 32 users, but NVIDIA also charges $100 for perpetual PC user licenses. The total cost of supporting 64 users, including licensing, on NVIDIA M10 cards runs roughly $11,400. That's $3,000 less than it costs to support the same number of users on NVIDIA's M60 GPU cards. AMD's MxGPU is still cheaper, but NVIDIA has a more proven track record. Finally, there is competition over midmarket GPU virtualization.

Now, Intel is making its first appearance. Intel Iris is a name long associated with graphics, but for the first time a hypervisor can use Intel's new GVT-g virtualization technology to slice that GPU up for multiple virtual machines. If you have a server with a Broadwell or Haswell processor, you can probably take advantage of it today -- well, if you also run XenServer 7. Granted, Intel Iris Pro Graphics is probably not the GPU you're looking for if you're trying to use AutoCAD or medical imaging applications in a virtual desktop. But the enterprise already has products that can address that use case. IT needs something for the 95% of users that would benefit from a bit of vGPU, and that's where this comes in.

There are limitations to the Intel Iris Pro Graphics platform right now, but it's early. The only device Intel showed at Synergy 2016 was an HPE Moonshot cartridge that could support seven users per CPU -- not per core, for the entire chip. Hopefully that scalability will improve. Now that IT has the high and midrange vGPU use cases covered, perhaps Intel can address the low-end users and allow every worker to have some sort of vGPU in their virtual desktops.

Next Steps

Everything you need to know about GPU virtualization

Hands-on look at how NVIDIA GRID works

Quiz: How well do you know vGPU technology?

See if your cloud workloads will benefit from GPUs

Dig Deeper on Virtual desktop infrastructure and architecture