It’s hard to believe, but this week’s NVIDIA showcase, the GPU Technology Conference (GTU), might be the biggest VDI conference of the year. This year, there are forty-nine sessions covering GPU virtualization for desktop workloads. Yes, that’s right: VDI is as big at GTC as it was at both Citrix Synergy and VMworld last year.
So, why the big focus on 3D VDI? With vSphere 6.0 shipping, vGPU support has transformed overnight from a technology that, while ready for mainstream adoption, was held back by the limited availability in vSphere 5, leaving it to Citrix to lead the way and limiting adoption to only those organizations willing to accept the extra management overhead of running multiple hypervisors. Now, however, with VMware on board, it opens the door to widespread adoption for all.
Not that Citrix will mind having the competition. The professional graphics market it still largely untapped, so this means net new license sales in a market that is less cost-sensitive than most. The cost of a XenDesktop or Horizon license is a far smaller part of the overall cost when you start loading servers with $3K GRID K2 cards. No matter what VMware might say, Citrix still retains a significant lead in desktop virtualization, and the cost of a XenDesktop license isn’t going to matter one bit if it means even marginally better performance.
Inevitably, Citrix and VMware are here in force, with what should be standout sessions from Citrix’s Derek Thorslund and VMware’s Mark Margevicius. All the big hardware vendors are here as well. Cisco is promoting UCS as the best platform for deploying graphics-intensive VDI. HP is promoting its engineering VDI stack (eVDI), which is based on its WS460c workstation blade, as well as its RGS protocol, which still has a loyal following in engineering shops. Dell is hosting what promises to be a sound introduction to vGPUs and workstation virtualization. It’s interesting to compare Cisco’s and HP’s approaches here. Whereas HP is talking workstations, Cisco is one of the few vendors that is actively promoting vGPUs for knowledge worker virtual desktops. Here, I think Cisco is on the right track. While there’s a big market for workstation-class VDI, the performance benefits of putting three or four GRID K1 cards in a server are significant, not only improving graphics performance but also allowing more virtual desktops per server by offloading all graphics processing from the CPU. This makes servers with four or more PCIe slots all the more important. The converged infrastructure appliance vendors, however, are nowhere to be seen. After competing so hard in 2012 and 2013 to deliver easy-to-use scale-out VDI appliances, they have done little so far to compete in the high-performance 3D virtual desktop market. If you find a VDI appliance with the space to accommodate more than one double-wide graphics card, you are doing well, and a multiprocessor appliance with space for only one GPU card is likely to be badly imbalanced.
From a system sizing perspective, GPU virtualization is significantly harder to get right than mainstream desktop virtualization. While it’s possible to get away without investing in third-party resource monitoring tools in many VDI environments, taking that shortcut with vGPUs in the mix is asking for trouble. With that in mind, Lakeside Software is appearing in two sessions, a clear indication of the importance that NVIDIA attaches to scaling professional graphics implementations correctly. Florian Becker’s Best Practices for Designing and Monitoring (S5111) session could well be the most valuable for engineers and architects already familiar with VDI and looking to incorporate GPU support for the first time. Also of interest is a session by Emily Apsey, from GIS software vendor Esri, who will be looking at performance-testing GPU-enabled apps in virtualized desktops, as well as a session by NVIDIA engineers Manvender Rawat and Jason Lee that covers scalability testing. We even have Pure Storage promoting the benefits of its platform as the storage element of a professional graphics VDI stack. I don’t know quite what Pure will be offering here, so I’ll hedge a little and say that while all-flash arrays deliver outstanding performance and are just about at the point where they make sense for general-purpose primary storage, AFAs are not the pinnacle of IOPS delivery for VDI. That honor remains with Atlantis Computing.
It’s not all just vendor sessions. The conference is packed with real-world examples of 3D graphics on VDI live in production today, with sessions from the likes of Bell Helicopter, PSA Peugeot Citroen, Ford, TATA, Huawei, Textron, and Dassault. And it’s not just engineering uses cases, either; Metro Health is showing how it uses Teradici APEX CPU offload cards and NVIDIA K1 cards to improve graphics performance in healthcare, and the team behind the BBC’s Doctor Who special effects will be here doing a live demo of how they use GPU-powered VDI to create the show’s special effects (no TARDIS required)—can’t miss that one.
I’ll provide more coverage of any interesting announcements later in the week.