vGPU and GPU passthrough for NVIDIA GRID K2 (v340.57) K160Q, K180Q and K280Q!
Citrix introduced architectural changes with XenDesktop 7.6 that on-going, will allow Citrix XenServer and/or NVIDIA GPU/vGPU enhancements to be automatically available to XenDesktop. Previously, the introduction of new vGPU types or GPU cards from NVIDIA had required a XenServer hotfix e.g. to use the K120Q and K220Q vGPU types designed for scalability back in March.
Last week saw the release by NVIDIA of two new vGPU types (K180Q and K280Q) at the opposite extreme, designed for maximum performance; whereby a vGPU type effectively represents an entire physical GPU. Assigning an entire GPU to a VM was an option previously only available via GPU-passthrough.
This step heralds a new era whereby a whole GPU can be managed via the flexibility of vGPU. Long term, the generic Citrix GPU management stack and GPU vendors implementing similar will open the way for vendors like VMware and Citrix to support vMotion and XenMotion (live migration) of GPU enabled VMs, should NVIDIA or other GPU vendors roadmap and implement such features. Allowing vGPU and entire GPUs to be managed by the same orchestration tools and using one set of drivers was intrinsically designed into the Citrix data models, the XenServer team have documented these models in beautiful detail and you can read more, here.
At the moment NVIDIA do not as yet support CUDA for vGPU and there are differences such as this between the vGPU and GPU GRID drivers. So some XenDesktop users will still want to use GPU pass-through or the equivalent VMware vDGA, if they are using heavily CUDA optimised applications.
At Citrix the unified architecture of XenApp and XenDesktop means like NVIDIA we are increasingly moving towards a broad continuum of GPU accelerated solutions and the release of the K180Q and K280Q is one more step in this successful joint engineering alliance broadening the options available to HDX 3D Pro consumers.
You can read the full details of the NVIDIA release as well as details of another new vGPU, the K160Q, in NVIDIA’s GRID documentation, detailing driver releases versions 340.57, 341.08.
If you have any feedback or questions please explore the Citrix forums, here; or the NVIDIA forums, here; where you can discuss your plans and experiences with other users, and our support and engineering staff.