Radeon and Tesla in one PC

Recently I’ve purchased Tesla P4 to kick off some ML. I have to admit it’s one cheap little monster card :slight_smile:
Simple 80mm fan can keep it under cooking temp easily and works like a charm for small scale machine learning.
I have unlocked all features by installing grid driver on it and it works amazingly well:

There is one tiny issue - Nvidia panel is saying that windows os is in charge of assinging GPU to tasks but when i open win10 settings there is no Nvidia gpu to choose from at all:

image

i can still use it if program is allowing me to pick Tesla from list like 3dmark but if there is no option default is always what windows choose.

Please see my Tesla mega thread

3 Likes

Will do thanks mate.

If you run into any snags post there someone is always willing to help, good luck.

1 Like

Win11 lets you choose the card individually, this might be an option.

If you don´t want the card for gaming (you have the AMD already…so…) and just for computing on CUDA, you actually dont need to switch anyhting, since the Card wont be adressed by the windows driver, but via CUDA:

May I ask what exactly you want to use it for?

I had similar scenarios with M40s and AMD Cards. Tampering with the registry like in the linked thread sometimes worked, sometimes not… Deleting all drivers and installing the headless GPU first (either remote or with the basic display adapter working) worked better in my cases… With WIN11, its super easy…

See my post here should help get you going
Below are screenshots of my Tesla P40 set up the same way same driver.
Edit: See this post for different ways to set up with different type display out cards.


Screenshot_38

I have read whole suggested mega thread and all I did was to switch EnableMsHybrid=2 to make it appear as low power card. Than I was able to assign this card to specific tasks in windows and it works perfectly fine.

I have used cover to keep the air tunnels, cut in half than bend it to hold fan (i have used 70mm fan from some sff dell PSU. It was only 2 pin fan but relatively thin and kindof quiet so wasn’t a problem. Now my tesla runs 26degr in idle and like 65 max in most demanding tasks. Since it’s fully unlocked I have started to use it for encoding video and ML. Even though people don’t recommend it I was able to stream on OBS on P4 playing on my Radeon - madness but footage looks crisp.

Yes, when using the grid driver it is a little less needing to be done to get it working as a pick as High Performance card but that can depend a good deal on what is being used as the display out card or passthrough card as some call it. I’ve even had the grid driver set things correctly without needing any mods but then again it came down to the passthrough card and driver version used.
Screenshot_62

1 Like

Yeah the Radeon cards tend to be a little more of pain to use for passthrough over Nvidia and Nvidia or igpu but it does with with Radeon just can be a little more fussy

Hmm, just about every AMD card I’ve tried has had the reset bug. My old Vega, my new 6700, I think even my old Fury pulled that crap. The VM will boot once fine, but then the whole damned host needs a reboot to reset the stupid card. Why is this still a thing? Is this AMD’s answer to Nvidia’s Code 43 error?

And that is why you buy the Nvidia card cause they have actuall support for VM passthrough and really doesn’t like at all they do dedicated comoute fine but VM stuff not so much vgpu is actually supported on geforce now

Fortunately I did not sell my 2070 Super. So it got transplanted into my “new” ESXi host. Works fine, no Code 43 error in sight. From what I can gather, Code 43 was an unintended problem, but Nvidia made zero effort to fix it. They probably would have much preferred customers using virtualization buy a Quadro, Tesla or Grid for that purpose. I’m not sure why they changed their tune, or perhaps the fix was as unintended as the original problem.

Oh no the code 43 was intended from Nvidia on geforce to stop people using them and having to buy workstation cards it was just a driver switch that’s all it is like the FP performance on matched dies compared to the geforce and Quadro thats why you can trick the drivers using a Quadro for display out but using a much better geforce for compute to get the Quadro features and better fp performance

1 Like

In that case, why did Nvidia change their position? It was usually easy to defeat once you knew how, but it was still…consumer unfriendly. I can’t say the same for the reset bug; I’ve been unable to get most AMD cards to work with GPU pass through even after supposed patches and workarounds have been applied.

For amd workstation card your better bios flashing a fire pro or Radeon bios onto the instinct if there is a matched card

1 Like

Radeon 6000 doesn’t have the reset bug

On ESXi, if I pass through my 6700 to a VM, I can only boot the guest once. If I shut the VM down or reboot it, then I must restart the ESXi host. If I don’t, then the VM will fail to boot and eventually it’ll time out with an error.

Oh Wendel hasn’t had they issues with rdna2

Wendell’s probably using KVM. Maybe it has some mitigations not present in ESXi?

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