Nvidia Grid is for sharing virtual GPUs (vGPUs) across multiple virtual desktop and applications with GPU instances.
List of Releases with respect to NVIDIA Grid.
• NVIDIA GRID August 2015 (1.0) with Nvidia K1 & K2 Card
• NVIDIA GRID April 2016 (2.0) with Nvidia Tesla M60 & M6 Server Card
• NVIDIA GRID May 2016 (3.0) with Nvidia Tesla M60 & M6 Server Card
• NVIDIA GRID August 2016 (4.0) with Nvidia Tesla M60 & M6 Server Card
• NVIDIA GRID August 2017 (5.0) with Nvidia Tesla P40 & P4 Server Card
NVIDIA has software to manage and monitor GPUs as a component of the data center called NVIDIA Quadro Virtual Data Center Workstation Software.
Quadro Virtual Data Center Workstation
The previous version with M60, GRID contained quadro profiles which ended in a letter Q.
With Tesla P40, Quadro vDWS profiles designed for workstations and high-end professional applications like Catia, Auto desk Maya, Siemens NX profession series capability. The floating licensing model Autodesk AutoCAD, Revit, Civil 3D, Adobe Creative Cloud 2015, and MATLAB works best. The new cards that support GRID 5.0 are the Tesla P4, P6, P40, and P100.
P40 single GPU board with 24GB Frame Buffer considered as the successor of M60. P40 can handle up to 24 users per board which means 1GB of Frame buffer per VM
M60 Dual GPU board with max 16 users per M60 card with 1GB frame buffer and 32 users with 512Mb of framebuffer. Now in Tesla P40, 512MB of framebuffer not available in GRID 5.0
|NVIDIA GRID P40 vGPU profiles|
|Name||Max Instances||FB Memory||Display Heads||Max X Res||Max Y Res||License|
|P40-1Q||24||1024MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
|P40- 2Q||12||2048MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
|P40- 3Q||8||3072MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
|P40- 4Q||6||4096MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
|P40- 6Q||4||6144MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
|P40- 8Q||3||8192MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
|P40- 12Q||2||12288MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
|P40- 24Q||1||24576MiB||4||4096||2160||GRID-Virtual-WS,2.0 & GRID-Virtual-WS-Ext, 2.0|
1. Changing configuration
Previously GPU Mode Switch utility to convert the Nvidia M60 GPU board from compute mode to graphics mode. It is also required to disable ECC memory.
In P60, there is no switch utility required. Built in NVIDIA-SMI utility is available so that no need to disable ECC memory
2. CUDA support
In M60 16GB card, there are 2 GPU board with each 8GB Frame buffer. In that case CUDA support was only for 8Q available. Any lesser 4Q, 2Q, 1Q will not get CUDA Support
In P40 24GB Card with max 24Users so each frame buffer size can have CUDA support(refer IMAGE 1)
3. Preemption support
Preemption support on Pascal cards allows all VMs to share GPU resources on round-robin time slicing fashion. Time allocated for all VMs are not same. Some jobs can take hours to complete. The CUDA jobs could fail when the time slice expired if preempt the running jobs is not available. So GRID uses round robin time slicing to share the GPU across multiple VMs
4. Fixed Share Scheduler
In M60 Graphics card, Fixed Share Scheduler not available.
In Tesla P40 card, each VM that is placed on the GPU an equal share of resources. Time slices are still used in the fixed share scheduler. As VMs are placed onto, or removed from, a GPU, the share of resources available to each VM is recalculated, and shares are redistributed to ensure that all VMs get equal access.
On per-application monitoring basis, Administrators can view GPU utilization using the NVIDIA-SMI tool.
Licensing is forced with GRID, and if a license is not available, the GRID drivers will not function.