NVIDIA
Questions about moving from Amazone to owned vmware infra
We have developed an application to do 3D rendering that is currently running under Ubuntu Linux V14 using SiftGPU http://cs.unc.edu/~ccwu/siftgpu/ with a K520. We want to move the code to a VMware infra maintaining the Ubuntu OS SiftGPU but seems we would need to get either a K1 or a K2 as the drivers for ESXi are only supported with those card. Questions: Do we have to use either the K1 or K2 or could we continue to use the K520. In either case what drivers do we need to use. Lastly, if we go with the K1 or K2 can our application running in multiple instances take advantage of the shared vGPU grid under VMware ? Thank you for your help and advice Jean D.
We have developed an application to do 3D rendering that is currently running under Ubuntu Linux V14 using SiftGPU http://cs.unc.edu/~ccwu/siftgpu/ with a K520.

We want to move the code to a VMware infra maintaining the Ubuntu OS SiftGPU but seems we would need to get either a K1 or a K2 as the drivers for ESXi are only supported with those card.

Questions:

Do we have to use either the K1 or K2 or could we continue to use the K520.
In either case what drivers do we need to use.
Lastly, if we go with the K1 or K2 can our application running in multiple instances take advantage of the shared vGPU grid under VMware ?

Thank you for your help and advice

Jean D.

#1
Posted 04/01/2015 07:42 PM   
[quote="John"] Questions: Do we have to use either the K1 or K2 or could we continue to use the K520. [/quote] You'll need to use the K1 or K2. K2 equates to the K520 and can leverage the same SDK. [quote="John"] In either case what drivers do we need to use. [/quote] I'd recommend using the latest released drivers, and selecting the appropriate drivers for your hypervisor implementation. Note that currently, vGPU does not have a Linux guest driver so you would need to use Passthrough. Current Linux drivers for Passthrough are here http://www.nvidia.co.uk/download/driverResults.aspx/82298/en-uk [quote="John"] Lastly, if we go with the K1 or K2 can our application running in multiple instances take advantage of the shared vGPU grid under VMware ? [/quote] Unfortunately, there is no vGPU driver for Linux guests. However it is under consideration as we see a slow growth in demand for Linux vGPU sessions. Also, if you're using the CUDA implementation of SiftGPU then you'll need to use Passthrough as CUDA is not currently supported on vGPU regardless of the Guest OS. I hope this helps clarify the options. You can absolutely migrate to a privately owned environment and deliver the same level of performance (if not greater) that you're getting in the Amazon cloud.
John said:

Questions:

Do we have to use either the K1 or K2 or could we continue to use the K520.


You'll need to use the K1 or K2. K2 equates to the K520 and can leverage the same SDK.

John said:
In either case what drivers do we need to use.


I'd recommend using the latest released drivers, and selecting the appropriate drivers for your hypervisor implementation.
Note that currently, vGPU does not have a Linux guest driver so you would need to use Passthrough.
Current Linux drivers for Passthrough are here


http://www.nvidia.co.uk/download/driverResults.aspx/82298/en-uk


John said:

Lastly, if we go with the K1 or K2 can our application running in multiple instances take advantage of the shared vGPU grid under VMware ?



Unfortunately, there is no vGPU driver for Linux guests. However it is under consideration as we see a slow growth in demand for Linux vGPU sessions.

Also, if you're using the CUDA implementation of SiftGPU then you'll need to use Passthrough as CUDA is not currently supported on vGPU regardless of the Guest OS.


I hope this helps clarify the options. You can absolutely migrate to a privately owned environment and deliver the same level of performance (if not greater) that you're getting in the Amazon cloud.

Jason Southern, Regional Lead for ProVis Sales - EMEA: NVIDIA Ltd.

#2
Posted 04/05/2015 09:04 AM   
Hi Jason, We are updating our project to move from Amazon. We are looking at Nutanix for the hardware platform along with Acropolis 4.5 (KVM) as the hypervisor, is this supported using the Linux version of the driver in a Virtual Machine in Passthrough mode ? Also any development on the " consideration for Linux vGPU sessions." ? For licensing, do i have the right understanding that passthrough mode does not requires special licensing "ex: Workstation extended" to use the CUDA implementation of SiftGPU ? Thx John
Hi Jason,

We are updating our project to move from Amazon.

We are looking at Nutanix for the hardware platform along with Acropolis 4.5 (KVM) as the hypervisor, is this supported using the Linux version of the driver in a Virtual Machine in Passthrough mode ?

Also any development on the " consideration for Linux vGPU sessions." ?

For licensing, do i have the right understanding that passthrough mode does not requires special licensing "ex: Workstation extended" to use the CUDA implementation of SiftGPU ?

Thx

John

#3
Posted 12/08/2015 05:02 PM   
[quote="John"] We are looking at Nutanix for the hardware platform along with Acropolis 4.5 (KVM) as the hypervisor, is this supported using the Linux version of the driver in a Virtual Machine in Passthrough mode ? [/quote] You'll need to speak with Nutanix regarding their support of the functionality. Currently we formally support RHEL KVM. [quote="John"] Also any development on the " consideration for Linux vGPU sessions." ? [/quote] It's a feature of GRID 2.0 which was released 31st August 2015. Virtual Workstation and Virtual Workstation Extended both include vGPU for Linux. [quote="John"] For licensing, do i have the right understanding that passthrough mode does not requires special licensing "ex: Workstation extended" to use the CUDA implementation of SiftGPU ? [/quote] For GRID 2.0 (M60 & M6) Passthrough requires a Virtual Workstation Extended license.
John said:
We are looking at Nutanix for the hardware platform along with Acropolis 4.5 (KVM) as the hypervisor, is this supported using the Linux version of the driver in a Virtual Machine in Passthrough mode ?


You'll need to speak with Nutanix regarding their support of the functionality. Currently we formally support RHEL KVM.

John said:
Also any development on the " consideration for Linux vGPU sessions." ?


It's a feature of GRID 2.0 which was released 31st August 2015.
Virtual Workstation and Virtual Workstation Extended both include vGPU for Linux.

John said:
For licensing, do i have the right understanding that passthrough mode does not requires special licensing "ex: Workstation extended" to use the CUDA implementation of SiftGPU ?


For GRID 2.0 (M60 & M6) Passthrough requires a Virtual Workstation Extended license.

Jason Southern, Regional Lead for ProVis Sales - EMEA: NVIDIA Ltd.

#4
Posted 12/09/2015 03:08 PM   
Scroll To Top

Add Reply