Tesla K40 questions

Hardware Enterprise Gear
1

I’m going to buy a K40 Tesla for blender, they’re only 100$ and that’s a lot of power and vram

Yes yes I know it’s a passive card that needs lots of forced airflow, I’m going to have that covered

My questions are about using it as a virtual GPU and passing through it’s output to another GPU like intergrated

I actually have a z68 board that supported the old Lucid Virtu software wondering if there is anything like that in the wild, heard Nvidia doesn’t like virtualization

Anybody have experiences or info or tips
Using it as a GPU would purely be a bonus my main use is for blender

2

Not seeing a lot about that online. I wonder what forum would have a quantity of info about that software.

I have used NVIDIA GPU’s for virtualization and RemoteFX (which was very very good), and support was seamless. So hopefully no issues there.

What host OS are you planning to use?

3

Like I only want passthrough to intergrated or to a 1050ti not “7 gamers one CPU” type stuff

4

I’m not sure what you mean. In the past when I did this, I could assign as many GPU’s to a guest OS as I pleased and they would be scaled appropriately as long as the NVIDIA WDDM driver was present in the host OS.

So that compatibility for whatever host OS you plan to use has to be there. In the case where I was using it, I was demoing Arcgis using RemoteFX and NVIDIA integration. That was excellent, Arcgis is a dog without all the power you can throw at it.

In that world, it’s the combination of OS/driver support that makes or breaks the experience.

On the other hand, I’ve seen Wendell do other tricks with pass-through were explicit driver support was not (or at least did not seem to be) key to these things. Activating the WDDM driver in Windows for RemoteFX, for example, required an extra license which was very expensive.

5

Basically I want to play games that render on the k40 and then pass through the output to another GPU since the k40 doesn’t have any

Yeah I know k40 is not a gaming GPU

But at the same time it does support all the gaming apis

6

Oh, yea that sounds b!tchen.

Hrm, that’s tough… You would probably have to do that as a virtualized networked vm. That’s how I would do it.

You’d also need a 10g nic (I would think, at least), on both sides on an isolated network off the internet.

I just want to add how surprised I am that that thing is passively cooled. That sounds crazy man.

7

Well it’s passive as in it doesn’t have fans on it, it relies on forced air from a servers chassis fans to cool it, I’m going to hook up a blower or delta fan to it

I think the reason they’re so cheap is that they don’t support Nvidia vGPU or grid

But like this thing is basically a Titan which isn’t anything to sneeze at at 100$

8

Nice, if you get it to work, you should publish it.

I’ve done this with Windows Server for Arcgis, but I’d love to see someone do it on something else.

9

Oh wait, so if it doesn’t support vGPU then I would think it wouldn’t work… Sounds like what you’ve got there is a card that’s built purely for sending and receiving various calc api calls. I’ve never researched cards of this type because they are far out of my typical budget.

In the past, I used Quadro cards, which also didn’t explicitly support vGPU, but had display port outputs so they were geared for rendering video.

I don’t know why they would even make a card that’s been whacked like that though… makes no sense?

10

Well I think vGPU is the new licensing thing they’re trying to push, where before they weren’t making money off it

11

Oh, that’s terrible. Yea, I dunno. I’ve been wanting to convert to a non Windows Server dependent system for a while because I’m fearful of this.

I wanted to implement Wendells passthrough idea for a long time, because it seems free of this sort of thing, but I don’t know what to do with all my existing Storage Spaces stuff. The integration is just soooo niiiiice though, ugh.

Really sad to hear NVIDIA is causing shenanigans with vGPU now. That’s a bummer, for sure.

12

Hey Giga, did you ever attempt hooking the K40 up and passing it through a GPU with video output?
I’m looking to do something similar, I have a quadro k2200 and was looking at getting a Tesla K20 to hook it up to. I found some information on Nvidia Maximus from what I can tell is just a special driver that facilitated the K20 to pass through the quadro.
If you have any experience with the Tesla I’d love to hear about it.

13

I did, I got very very close using an 2400G on a AB350N wifi, but on my AX370 Gaming 5 it appears 4G decoding is somewhat broken and couldn’t handle all the BAR space of both a 8GB vega 64 and a 12GB tesla K40

it could also be that I am an idiot, I am linux noob and I basically had to be told command by command how to do it

14

I’m looking to do something similar. Is there any way I can use two k40s as Cuda compute devices? I have a b350 + 1600x rig so I’m not sure if motherboard supports it.

15

you absolutely can use them as cuda compute devices, just not in a VM

you’ll need above 4g decoding turned on as well as CSM disabled, you need to install windows on a disk formated with GPT

now my gigabyte gaming 5 (not k) x370 did not like my K40 with vega64 and I believe this is due to gigabytes lackluster above 4g decoding support and taking up too much bar space

if you have any more questions or need help please make a new thread as anymore will be necro posting

16

Those cards [Tesla series] have like 90ish % hardware / software, but not all necessary finite bits, for allowing rasterization / pass through. Some small-AF bits can convert it, to a Grid series GPU equiv. --> HardMod/Soldering req’d + BIOS flashing [best of my knowledge, works for K10 -> Grid K2 conversion]

17

This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.