Grid newbee, want to add CAD VDI to serveur. Need help

Hi gang,

I’m new to the forum and to Grid servers.
At the moment I have 20 Autocad heavy users using workstations w/mid-level pro vga cards.
I would like to move these users to VDI.
Can anyone tell me how many Tesla M60s I need to give them equal or better performances?

Thanks everyone,

the M60 1Q profile should be sufficient. Therefore you get 16 users on one M60. But be aware that you also need CPUs with high clock frequency >3Ghz especially for Autodesk products.
So for redundancy I would at least recommend 2M60s for your deployment


Thanks for the info. Is there somewhere where I can find a list of the Tesla M cards that mentions how many users at what level of performance each cards provide? Ex: Tesla M60 provide Quadro M4000 performances to 16 users.

Hey gang,

Got this server Dell R730, and I’m wondering how powerful would 5 VDI VMs be?

Server Description:
PowerEdge R730 Server
Shipping and Handling
Dell EMC 2U Standard Bezel
PE R730/xd Motherboard MLK
R730/xd PCIe Riser 2, Center
R730 PCIe Riser 3, Left
R730/xd PCIe Riser 1, Right
Intel Xeon E5-2667 v4 3.2GHz,25M Cache,9.60GT/s QPI,Turbo,HT,8C/16T
(135W) Max Mem 2400MHz
Intel Xeon E5-2667 v4 3.2GHz,25M Cache,9.60GT/s QPI,Turbo,HT,8C/16T
(135W) Max Mem 2400MHz
PowerEdge R730 Shipping
Chassis with up to 16, 2.5 Hard Drives
Advanced ECC
DIMM Blanks for System with 2 Processors
16GB RDIMM, 2400MT/s, Dual Rank, x8 Data Width x 8
2400MT/s RDIMMs
Heatsink for GPUs PowerEdge R730
Heatsink for GPUs PowerEdge R730
Performance BIOS Setting
iDRAC8 Enterprise, integrated Dell Remote Access Controller, Enterprise
1.6TB Solid State Drive SAS Write Intensive MLC 2.5in Hot-plug Drive,
PX05SM x 8
PERC H730P Integrated RAID Controller, 2GB Cache
No Media Required
NEMA 5-15P to C13 Wall Plug, 125 Volt, 15 AMP, 10 Feet (3m), Power Cord,
North America x 2
Dual, Hot-plug, Redundant Power Supply (1+1), 1100W
No Trusted Platform Module
R730 GPU Installation Kit
NVIDIA GPU Power Cable
Broadcom 5720 QP 1Gb Network Daughter Card
Dell Optical Mouse MS116 - Black
Dell KB216 Wired Multi-Media Keyboard English Black
No Operating System
No Systems Documentation, No OpenManage DVD Kit
OpenManage Essentials, Server Configuration Management
Windows Server 2016 Standard Edition,Secondary OS, No MEDIA, 16 CORE
5-pack of Windows Server 2016,2012 USER CALs (Standard or Datacenter)
Windows Server 2016 Standard,16CORE,Media Kit
ReadyRails Sliding Rails Without Cable Management Arm
Unconfigured RAID for H330/H730/H730P (1-16 HDDs or SSDs)
UEFI BIOS Boot Mode with GPT Partition
On-Site Installation Declined
Declined Remote Consulting Service
MISSION CRITICAL PACKAGE: Enhanced Services, 5 Year



FYI - Each M60 GPU is "roughly" that of an M5000 …

The above server spec is fine. Looks like there’s only 1x M60 which will be your limiting factor for profiles and with 5x users (an odd number on a GPU with “even” GPUs and high Frame Buffer allocation) makes it slightly more tricky when dealing with “heavy” users.

You’ll need to define what you mean by “heavy” users. Heavy in terms of what? CPU? GPU? … As this will shape your Workstation resource profiles.

As I know absolutely nothing about your workload, current performance issues, data sizes, how many monitors you’re using, monitor resolutions or your performance expectations etc etc, for “heavy” users, you’re going to be looking at something like the following:

2x Workstations with the following:

8x vCPU

3x Workstations with the following:

4x vCPU

Obviously, with an “odd” number of users and a server (more relatively the GPU) that has an “equal” number of GPUs and higher frame buffer profiles, it’s not possible to give all users the same “Heavy” Workstation specifications. You’ll have to test and monitor utilization and workout what’s best for you.

If you had 4x users, I’d say do this for all profiles:

8x vCPU

Or if you want all specifications to be exactly the same, with 5x users you could do something like:

6x vCPUs

Play with the different configurations and see what works for your users. Don’t forget you’re running the newest v4 CPUs and newer GPUs, so you may need less resources to deliver a better experience, but as you’ve not mentioned the current Workstation specifications, or what kind (if any) of performance problems your users are currently experiencing, I don’t know.

Don’t forget to factor in that there will be more running than just AutoCAD. Email, browsers with multiple tabs and content. Maybe another AutoDesk product etc etc

The more information you can give us, the more we can help you.



Thanks for your reply Ben.

The workstations I want to replace with this server are as follows:
Xeon e5-1650 v3 @3.5GHz
AMD FirePro W5100 4GB
Gigabyte Network
SSD Data server
M.2 950 local OS drive

They run Autocad, Adobe CS, Rhino, Sketchup, Office, Google Chrome, Maxwell Render and V-ray.
The Autocad projects are very large with many objets and many very large xref.
Example of projects: Football Stadium

That should be fine, but your limiting factor is still frame buffer. I would ask if you’ve measured frame buffer usage on the AMD GPUs, but 4GB really isn’t a lot for a CAD user, and the only sensible option below that would be 2GB, which looking at those workstation specs and application list, is a non starter.

Which version of Windows are your CAD users running? How many monitors and what resolution?

If all 5 users do need 4GB of FB, you’ll need to purchase a second M60, which will be the cheapest but still expensive option as it will be hugely under utilized.

Making the whole solution cost in (not just GRID) is typically about density, the more users you can get on the system (without impacting performance), the cheaper it is, the fewer users, the more expensive. Having all high performance users on a single server is not the best way to do it, and it is going to be expensive due to the lack of density. Ideally you mix the high performance with low / medium performance users and effectively “plug the gaps” in terms of resource allocation. Not forgetting that whatever physical resource you have, you need N+1 for resilience, unless you have some very understanding CAD users and a good explanation of why they’ve all been wiped out at the same time when an outage happens …

However, sometimes there are only high performance users and you have to plan accordingly. 5 is just a very unfortunate number in this case, and that 1 extra user is going to cost a lot to support. As said above, with a 4GB requirement, you’ll need to purchase a second M60 just to get all users connected, then a second R730 with identical spec if you want the N+1

FYI - The N+1 (and let’s be honest, it’s N+1 as an absolute minimum in any real production environment, typically you should have greater than that) resilience is the bit a lot of vendors conveniently miss out when quoting how cheap this kind of solution is.

What is it that’s driving the Workstation replacement? You mention “better performance” in your first post, but are there any other factors that are contributing to it?



Hello JLTrepanier, i see you choose the Xeon E5-2667 v4 3.2GH. Did you choose that one for the high 3.2GHz cpu for CAD purpose instead of for example a cheaper Xeon E5-2650 with 2,2GHz but with more cores?
I’m in process of assembling a similar HP Proliant server for the same purpose.
I mean, should we go for cores or for GHz?

Typically with CAD workloads, you’ll want Clock Speed over Cores. However workloads and data-sets can vary in requirements, so ensure you assess current limitations before ordering.

If you require more density than the E5 2667 v4, but a slightly lower level of performance, the E5 2697A v4 is a great choice.



The specs of the 2697A look very good but then we go over our budget :-)
The new Microsoft core licensing model would be a reason to go for more HZ instead of cores.
Thanks for the advice…

Yes, the new MS licensing is not great and does pose problems when you want to use a specific CPU.

I would only recommend the E5 2667 v4 for low density / High Performance use cases.



Hi Ben,

First, why 5 users? It was partly random, partly size of small architect firm.

By your recommandations above, this rack would be best for 4 users and would need to multiply it by groups of 4 users. Is that right?

I’m looking into this type of solution as a new business model.

That’s why 5 users :-D

Yes, absolutely, for high performance users with a 4GB FB requirement on 8GB GPUs, a multiple of 4 will give you the best overall utilization with an M60

Depending on how hard they use the CPUs, I’d be tempted to increase the RAM and drop another M60 in there to get better density. Don’t push it too far though, or you’ll reduce their performance.

If this is a new business model, then you’ll need density to make it cost effective. As mentioned above, this CPU: E5 2697A v4 is a great choice. It’s expensive, but it’s expensive for a reason. I’ve personally used it with some heavy 3d Seismic applications, and can vouch for it’s performance.

Anyway, regardless of any advice or pointers on here, monitoring the hardware when in use is absolutely key to density and performance. Then spec accordingly before continuing with the deployment. Don’t just order kit without testing first (regardless of spec), it may prove to be be an expensive mistake, I’ve seen it happen!!



Thank Ben for everything this is great help.

To say, I was wondering if I could mix in one or two GForce cards for rendering?

No problem.

Can you actually virtualise other GPUs, yes, you can. Virtualise as in “Passthrough”, so these are a 1-1 mapping only. There’s no GRID software to put on top to give you any additional features. However, you run into all sorts of other issues such as hardware compatibility and support with the server vendors, appropriate cooling of the GPUs, there’s no monitoring of the GPU except within the VM etc etc. It’s not a scalable solution. Plus it kind of negates the whole reason for virtualising anyway, which is to make better use of the hardware.

Also, neither GeForce or Quadro go through the same testing process as Tesla. Basically, Tesla GPUs are designed to run hard 24x7x365. GeForce and Quadro are not.



I was thinking of going Passthrough for the Gforce and giving it only to a Virtual Rending App, if it’s possible. It wouldn’t run 24/7, it would be just as needed. It would help me fill in the “gaps” and use the server better.

Thanks again helping me through this thought process.