Question Can I build a numerical modeling computing system that can also support nvidia vgpu and remote working at the same time ?

[antlevog]

I would like to build a super PC that can support two main functions:

1. Run MIKE21 (of DHI) numerical modelling software.
2. Support working remotely for at least 10 CAD users using nvidia Quadro's virtual GPU software

MIKE21 is using only Intel MPI and Nvidia GPUs so this is a restriction. Moreover MIKE21 does not support new CPU architecture that has Power and efficient cores. Interesting enough is the fact that it can officially run on windows server.

Having all the above in mind I had concluded to a Intel® Xeon® W9-3495X CPU, with two or three Nvidia Quadro A5000 Ada GPU.
It never occurred to me that it would be that much difficult to find such a piece of hardware compatible with Windows Server!

Can you guide me through the market to purchase the proper hardware / software (OS) for the above mentioned functions?
Thank you in advance.

Best Regards,

 

[Ralston18]

I believe that the requirement is actually a server vs "super PC" correct?

And the requirement is that the proposed system must be a Windows Server - correct?

What have you done or attempted with respect to finding potential sources or vendors?

What specifically is making the purchases difficult? What incompatibilities are being encountered?

I googled "Where to buy a Intel® Xeon® W9-3495X CPU" seems to be a number of potential sources as I understood some of the resulting links.

For example (not an endorsement or recommendation):

https://www.avadirect.com/Xeon-w9-3...-4677-420W-MTP-OEM-Processor/Product/15775925

There were other similar vendor links.

More information needed.

 

[USAFRet]

Support working remotely for at least 10 CAD users using nvidia Quadro's virtual GPU software

This is a "server", not a "super PC".

 

[slightnitpick]

I have no experience here, but based on what you and everyone else wrote I'd say try this link: https://www.nvidia.com/en-us/data-c...jZXJ0aWZpZWRGaWx0ZXJzIjp7fSwicGF5bG9hZCI6W119

I couldn't find any A5000 ADA systems that were vGPU qualified, but there are some A6000 ADA that are. Other than that, the Gigabyte machines seem to fit the bill.

Edit: I'd recommend getting an opinion from DHI as well. This is their expertise. https://support.dhigroup.com/

 

[kanewolf]

This is a "server", not a "super PC".

The software licenses for 10 CAD instances and the OS that supports GPU virtualization may be a significant percentage of your total cost.
It has been several years since I configured a Virtual desktop infrastructure, but at the time Xendesktop was THE choice for virtualized remote desktop -- it is not cheap.
You should also consider if 2 or more hosts makes more sense. If your "Super PC" fails then you are 100% down. If you have two hosts then you are only 50% degraded.
In addition, most virtual desktop software requires some backend hardware to manage Windows AD, database, etc. When I configured virtual desktops, we allocated dedicated hardware to backend duties.
You REALLY need to start with your software infrastructure stack. Determine ALL the requirements to do the virtual desktops. Then, look at hardware.

 

[antlevog]

Thank you all of your prompt responses. I will try and respond to each of you separately.

 

[antlevog]

I believe that the requirement is actually a server vs "super PC" correct?

And the requirement is that the proposed system must be a Windows Server - correct?

What have you done or attempted with respect to finding potential sources or vendors?

What specifically is making the purchases difficult? What incompatibilities are being encountered?

I googled "Where to buy a Intel® Xeon® W9-3495X CPU" seems to be a number of potential sources as I understood some of the resulting links.

For example (not an endorsement or recommendation):

https://www.avadirect.com/Xeon-w9-3...-4677-420W-MTP-OEM-Processor/Product/15775925

There were other similar vendor links.

More information needed.

1. Yes. Since it will support 10 users is actually a server. But from the point of view that (in case it has a w9-3495Χ -56 core workstation CPU) is a "Super PC". Am I worng? Is there a "Super PC" definition that I am missing? If that is the case, apologies. Is just a server.

2. Well for compatibility issues (Mike21, Autocad) there must be a windows operating system onboard. Taking that into consideration as well as the fact that I need the machine to support 10 different users simultaneously Windows Server seems my only option. I am also not an expert, so if you see other options here please advise.

3. Purchase difficulties are as follows:

- I live in Greece (Europe).
- Nobody supports a built to order option of that kind of a machine.
- Everyone (resellers) refers to Dell, HP or Lenovo (these are the companies I have reached so far through my resellers). However all of these companies cannot combine Xeon-w9-3495X with Windows Server. This -to my understanding- is due to driver incompatibilities between the Motherboards that support this CPU model with Windows Server. To put it in another way they claim that this CPU is for workstation (not server) use, hence they do not bother to issue mobo drivers for windows server.
- On the other hand, if I turn to a server Xeon CPU, I was told that this kind of CPUs, do not support Intel MPI (please confirm this), which is necessary for MIKE21 to work smoothly (according to DHI support), plus they do not have the clock speeds of a w9-3495X CPU. plus most of Server CPUs nowadays have "power / high priority" and "efficient/low priority" cores which does not work well in MIKE21 (this is confirmed both in practice and with DHI support)

The general concept is to utilize my hardware resources as much as possible.
So for numerical modelling (MIKE21) I need as much CPU (power) cores, at as high clock speed as possible. However we do not compile numerical models 24/7. So I thought that instead of buying a separate server for 10 users working remotely, with a separate GPU, why can't I use the "Beast" that I will have from MIKE21? Even if I "sacrifice 10 - 20 cores for all users (2 cores per user), there are still 36 cores for MIKE 21 to simulate! And when there are no remote users onboard, MIKE21 will be able to use all 56 cores!

This was my goal. Does anyone thinks I have a chance here or should I divide the two different uses (MIKE21 and Serving 10 users with CAD)?

 

[antlevog]

This is a "server", not a "super PC".

Point taken.

 

[antlevog]

I have no experience here, but based on what you and everyone else wrote I'd say try this link: https://www.nvidia.com/en-us/data-c...jZXJ0aWZpZWRGaWx0ZXJzIjp7fSwicGF5bG9hZCI6W119

I couldn't find any A5000 ADA systems that were vGPU qualified, but there are some A6000 ADA that are. Other than that, the Gigabyte machines seem to fit the bill.

Edit: I'd recommend getting an opinion from DHI as well. This is their expertise. https://support.dhigroup.com/

Thank you for your response and your link.

I had extensive talks with DHI, regarding their hardware requirements (at first I was aiming for an AMD threadripper but they do not support it due to the use of Intel MPI in their algorithms.)

Regarding this link (which is in fact interesting and thank you for that), I find my self in the following problem:

Regarding the Fujitsu PRIMERGY models: They do support windows Server BUT the supported CPUs (4th Generation Intel® Xeon® Scalable Processors ) have high and low priority cores. This feature is not supported by MIKE21. It confuses the algorithms and it plumets the overall performance of a simulation.

Regarding the Gigabyte Tower Server / Workstation, at first I was delighted (because of the "Server" term on a W-3400 machine) but then I noticed under the specification that Windows Server is not under the supported operating systems.

https://www.gigabyte.com/Enterprise/Tower-Server/W773-W80-rev-10#Specifications

 

[antlevog]

I am curious though. What kind of hardware is not supported on these motherboards from windows Server? I mean if this is about the LAN card, I can buy a compatible one and "kill" the onboard. Is that the case?

Then there is this (just discovered it): https://www.intel.com/content/www/us/en/support/articles/000093607/processors.html, which frankly i do not understand: How is it possible a CPU not to be supported by an operating system?

 

[kanewolf]

Having all the above in mind I had concluded to a Intel® Xeon® W9-3495X CPU, with two or three Nvidia Quadro A5000 Ada GPU.

Why a single CPU rather than a dual socket motherboard? This goes back to the server vs workstation differences. Server chassis typically are dual socket motherboards.

 

[Ralston18]

Regarding " How is it possible a CPU not to be supported by an operating system?"

It is a matter of purpose/requirements and design.

I have a pickup truck. I could modify it to go faster. I could modify it to carry a heavier load. I could modify it to float. I could modify it to use battery power. However, to meet all or even some of those requirements at the same time would be quite problematic.

= = = =

MIKE21:

This software?

https://www.dhigroup.com/technologies/mikepoweredbydhi/mike-21-3

https://support.dhigroup.com/knowledgebase/article/KA-01323/en-us

Per the second link Windows OS is supported.

= = = =

Noted: "The general concept is to utilize my hardware resources as much as possible."

The root of all this is, I believe, that the intent/requirement is to put together a system fully made up of optimized "components" (operating system, hardware, applications, performance, multi-tasking, remote use, virtual desktop, etc..

Fair enough in general....

Unfortunately what may be optimal for one requirement is not fully optimal for another requirement.

There will need to be trade-offs to achieve some sort of balance.

What are the priorities with respect to the end operational/functional requirements?

List and quantify the priorities and trade-offs you are willing to make.

And then there are considerations such as budget, purchasing, support, etc..

Those may be even more constrictive.

 

[antlevog]

Why a single CPU rather than a dual socket motherboard? This goes back to the server vs workstation differences. Server chassis typically are dual socket motherboards.

because of three things mainly:

1. Server CPUs do not support Intel MPI (at least that is what I have heard)
2. Server CPUs have in general lower Clock Speeds
3. Server CPUs have high and low priority cores. This is devastating for MIKE21 numerical modelling performance.

 

[antlevog]

Regarding " How is it possible a CPU not to be supported by an operating system?"

It is a matter of purpose/requirements and design.

I have a pickup truck. I could modify it to go faster. I could modify it to carry a heavier load. I could modify it to float. I could modify it to use battery power. However, to meet all or even some of those requirements at the same time would be quite problematic.

= = = =

MIKE21:

This software?

https://www.dhigroup.com/technologies/mikepoweredbydhi/mike-21-3

https://support.dhigroup.com/knowledgebase/article/KA-01323/en-us

Per the second link Windows OS is supported.

= = = =

Noted: "The general concept is to utilize my hardware resources as much as possible."

The root of all this is, I believe, that the intent/requirement is to put together a system fully made up of optimized "components" (operating system, hardware, applications, performance, multi-tasking, remote use, virtual desktop, etc..

Fair enough in general....

Unfortunately what may be optimal for one requirement is not fully optimal for another requirement.

There will need to be trade-offs to achieve some sort of balance.

What are the priorities with respect to the end operational/functional requirements?

List and quantify the priorities and trade-offs you are willing to make.

And then there are considerations such as budget, purchasing, support, etc..

Those may be even more constrictive.

I am afraid you are right in all your points. And yes MIKE21 is not only Windows Server compatible. It is also compatible with Windows Desktop (7, 10,11). As you pointed out the root of the "problem" is my intent to combine different uses (MiKE21 vs Working Remotely). I wanted to believe there would not be a necessity of compromising...It seems that I was wrong. This forum was my last hope in finding some middle ground.

Still, I need to confirm whether or not Server Intel CPUs do not support Intel MPI, in order to prioritise. Budget might not be such a concern, since I will pay 50% less of any purchase through a finance program.

 

[Ralston18]

Re: "I need to confirm whether or not Server Intel CPUs do not support Intel MPI".

Someone here may know and can post accordingly. I (full disclosure) do not know.

However, in the meantime, simply pose that question directly to Intel as specifically as you can.

In writing and thus be able to document the answer(s).

 

[kanewolf]

because of three things mainly:

1. Server CPUs do not support Intel MPI (at least that is what I have heard)
2. Server CPUs have in general lower Clock Speeds
3. Server CPUs have high and low priority cores. This is devastating for MIKE21 numerical modelling performance.

According to this page -- https://www.intel.com/content/www/u...irements/mpi-library-system-requirements.html Xeon Scalable CPUs are MPI library compatible.
Clock speeds depend on the specific CPU you choose.
You should read this -- https://community.intel.com/t5/Processors/Xeon-High-Low-Priority-Cores/td-p/1471916 about priority cores. They are not the same as P and E cores on desktop CPUs.
You need to research the Xeon Scalable CPUs more, IMO. Start at ark.intel.com

 

[slightnitpick]

https://www.thewindowsclub.com/enable-rdp-for-multiple-users-on-windows-11

It looks like it might be possible to modify Windows 11 Enterprise to all multiple remote desktop connections. I don't know if you want to go down this route, though, as you might have to buy everything first and then hope it works (though if you have another Windows 11 machine that you can test this on, that might be worth doing). It might be worth emailing Microsoft to see if they have a solution.

 

[kanewolf]

https://www.thewindowsclub.com/enable-rdp-for-multiple-users-on-windows-11

It looks like it might be possible to modify Windows 11 Enterprise to all multiple remote desktop connections. I don't know if you want to go down this route, though, as you might have to buy everything first and then hope it works (though if you have another Windows 11 machine that you can test this on, that might be worth doing). It might be worth emailing Microsoft to see if they have a solution.

Remote desktop is not GPU enabled with a Windows desktop OS. Not suitable for CAD instances.

 

[antlevog]

Remote desktop is not GPU enabled with a Windows desktop OS. Not suitable for CAD instances.

Thank you all for your answers. I haven't fully read all your links but it is enlightening! I will and revert.
Mean while let me share on more thought (and link) which occured to me when I saw this statement:

- Should I use VMware instead of Remote Desktop then?

- It seems that nvidia supports vGPU through VMware (see here:https://docs.omniverse.nvidia.com/deployment/latest/vgpu-virtual-desktop.html)

- Does this means that I can use VMware on a Windows 11 physical machine for working remotely cad software?

Please advise.
Thank you in advance.

 

[kanewolf]

Thank you all for your answers. I haven't fully read all your links but it is enlightening! I will and revert.
Mean while let me share on more thought (and link) which occured to me when I saw this statement:

- Should I use VMware instead of Remote Desktop then?

- It seems that nvidia supports vGPU through VMware (see here:https://docs.omniverse.nvidia.com/deployment/latest/vgpu-virtual-desktop.html)

- Does this means that I can use VMware on a Windows 11 physical machine for working remotely cad software?

Please advise.
Thank you in advance.

I don't know for sure, but my guess is that VMWare has to be the bare metal OS.

 

[antlevog]

I don't know for sure, but my guess is that VMWare has to be the bare metal OS.

I think you are right, and I think that this might be the solution;

Use VMWare as a hypervisor OS or Xendesktop or something relevant, and divide the Intel® Xeon® W9-3495X cores and the nvidia GPU cores in several virtual machines using Windows 11!

With such configuration we address the (in)compatibility issue of W9-3495X & Windows Server, plus we can use Nvidia vGPU creating several "remote workstations" for CAD remote working (Overcoming the RDP GPU issues).

There are two main issues that need to be assessed (if I am still thinking straight):

Whether or not VMware (or relevant hypervisor software) is compatible with W9-3495X and how much will be the cost of VMware subscription, along with Nvidia vGPU license subscription and if there are perpetual licenses for each software.

Do I miss something here? Is it too good to be true?
Apologies if I am talking nonsense here. Any advice will be valuable since I am new to this, and it is really hard to find a local expert that knows this field.

 

[kanewolf]

Whether or not VMware (or relevant hypervisor software) is compatible with W9-3495X and how much will be the cost of VMware subscription, along with Nvidia vGPU license subscription and if there are perpetual licenses for each software.

I don't believe there are perpetual licenses. You will be paying annual fees.