Virtualized GPUs have extended VDI's appeal by delivering high-end graphics to remote users, and broader ecosystem...
support will soon make the technology accessible to more IT shops.
A major theme of Nvidia Corp.'s annual GPU Technology Conference next week will be the expansion of its hypervisor ecosystem beyond Citrix XenServer.
That could mean Nvidia's Grid Kepler architecture with virtualized GPU (vGPU) technology will be supported on VMware Inc.'s vSphere. It may also be supported on Desktop as a Service platforms, including VMware Horizon DaaS and Amazon WorkSpaces, so cloud-hosted virtual desktops can run graphics-intensive apps just as well as a PC, said Will Wade, director of Grid products.
VDI: CPUs vs. vGPUs
Two years ago, Nvidia launched pass-through graphics support for power users that need a dedicated GPU. The company also delivered the vGPU, one of three technologies in its VGX platform, which allows multiple users to share a GPU, increasing server utilization.
The technology has become necessary for virtual desktop infrastructure (VDI), because most desktop environments today are graphic-centric.
"The use cases of 3-D graphics for virtual desktops are huge, and the adoption is increasing fast," said Ruben Spruijt, chief technology officer of PQR, an IT services provider based in the Netherlands.
Even Windows 7 requires that PCs include a DirectX 9 graphics device, and Microsoft recommends advanced graphics hardware for certain features, including video playback.
A move from PCs with graphic devices built in to virtual desktops where everything runs on a CPU could mean stodgy performance. A CPU can run graphic-intensive apps (software rendering), but it isn't designed for it and isn't efficient, experts say.
Ruben SpruijtCTO, PQR
"When you try to [run graphics] with VDI and CPUs, you get an experience you'd get in 1995," Wade said. "Can they be productive? Yes. But will they like it? No, and they can't function with apps like Photoshop, CAD, WebGL or video apps."
VDI administrators shouldn't throw vGPUs at every user, however. These virtual desktops cost about $50 to $100 more per user than a standard CPU-powered virtual desktop, just as PCs for power users cost much more than a typical users' PC, Wade said.
"[VGPU] isn't a commodity; we aren't designed for task workers," Wade said.
Up to eight users can be supported per physical GPU, depending on the vGPU profile, according to Nvidia's website. The Grid K1 boards pack four Kepler-based GPUs and 16 GB of memory, and host the maximum number of concurrent users. Grid K2 boards include two high-end Kepler GPUs and 8 GB of memory to maximize density for graphics-intensive apps.
Before deploying vGPUs, it's important to understand the user population, what type of applications they use and the performance and resource impact of the applications, Spruijt said.
A clear view on application resource usage includes GPU, CPU, IOPS, storage and networking, along with an understanding of VDI, Spruijt said.
"Does the application require a GPU to operate? Is the application GPU-intensive, or is the GPU just an assistant for the application?" said Spruijt, who will present a session on vGPUs next week at Nvidia's conference.
Nvidia integrates with monitoring tools such as Lakeside Software Inc.'s SysTrack to measure vGPU utilization. IT can use the tool to gather data for vGPU planning and monitoring to allocate resources efficiently.
Virtualized GPUs comparisons
While Nvidia leads in the virtualized graphics processing space, there is also Advanced Micro Devices Inc.'s (AMD) FirePro GPU. But Nvidia has an extensive list of certified servers from companies such as Cisco, Hewlett-Packard and Dell, and application certification and independent service provider support is another big differentiator, Spruijt said.
"AMD really needs to share their vision -- sell and market their upcoming solutions as soon as possible to win market share in the 3-D graphics for virtual desktops space," Spruijt said.
"3-D Graphics for Virtual Desktops Smackdown" white paper (.pdf)
An independent overview of vGPU technologies, co-written by Ruben Spruijt, Shawn Bass and Bernhard Tritsch, will be available early April on BrianMadden.com. The findings of the "3-D Graphics for Virtual Desktops Smackdown" white paper will also be presented at the Nvidia GPU conference, BriForum and Citrix Synergy.
On the software side, Citrix leads VMware and Microsoft with its pass-through support, vGPU and GPU-sharing support.
With Citrix, each XenDesktop VM gets its own dedicated vGPU driver, which sends commands and controls to one physical GPU using its own dedicated input channel. The driver returns rendered frames back to the virtual desktop, which then streamsit back to the remote host, according to a recent Nvidia Grid with VDI white paper.
For VMware users, Nvidia's graphics technology is interoperated through third-party tools such as Teradicis PCoIP.
It captures rendered Window images, encodes them via PCoIP and streams the Window content to the endpoint, according to the white paper.
VMware's pass-through GPU support for graphics users comes from its own virtual Dedicated Graphics Acceleration technology. It also has virtual Shared Graphics Acceleration (vSGA), similar to Nvidia vGPU. However, a comparison video by industry analyst Gunnar Berger shows that the vGPU in XenServer outperformed vSGA vSphere.