Tesla K80 Graphic Card
Petercat
Posts: 2,321
Would it work? Even if it renders slower than my current 980ti, my biggest problem with speed is scenes dumping to CPU, so it wouldn't matter if I could have the full 24GB RAM of the K80 available.
I'm using my 980ti for rendering only, on-board graphics to drive the display.
Win7.
Daz 3D is part of 
Connect
DAZ Productions, Inc.
7533 S Center View Ct #4664
West Jordan, UT 84084
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2025 Daz Productions Inc. All Rights Reserved.
Comments
I don't think so, that thing is meant for an entirely different sort of job. It is doubtful DS/iRay would have any idea what to do with that.
Your 980ti has, IIRC, 6Gb of VRAM so upgrading to a 2080ti would get you 11 Gb of VRAM or the latest Titan would get you 12. Those seem the best options if avoiding CPU rendering is your goal. Further the 20xx cards definitely can pool VRAM through NVLink so if you have the sort of money it takes to buy a K80 getting a second 20xx card shouldn't be out of the question for you if 11 Gb isn't sufficient for your needs.
Unless things have changed, it is a bad idea to mix Quadro/Tesla cards and GeForce (GTX/RTX) cards.
I wouldn't be mixing them, I have another spot for the 980ti.
So would it work?Never mind, found it, yes it will work.
I'll have to make a waterblock for it, though. Such a pain. Heh-heh.
Actually, I have no intention of going beyond the 1080 series for now.
Unless something drastic changes, but I have little hope of that.
Where did you find anything saying there is a driver for the K80 that iRay can talk to? I couldn't find anything about there being one anywhere.
Using both a Quadro and a GTX (mixing them) in the same machine can be done, but the effectiveness of the Quadro comes into question if the GTX drivers are loaded.
I'm running a reverse setup: Quadro K4000 for primary display and a 980 for rendering (more cores, more VRAM). MSI Afterburner doesn't see the Quadro, but it works just fine for primary display - faster screen refresh rate than the onboard mobo GPU, but then it's a K4000, so it's either that or a doorstop. I have an fx1800 for a doorstop already. If I have to use ot for an ice scraper, I'll upgrade my doorstop to the K4000.
The Tesla has CUDA cores, and a PCI interface. It merely lacks output jacks. It's my understanding these are typically used as slaves for rendering systems as well as number-crunching and all that fancy shmancy Scientific stuff like splitting proteins and folding protons and shtuff.
If your PC has multiple free PCI slots, you can get 2 K40s for less than a single K80 and come away with 5760 cores and still have 12GB of VRAM (24 if they pool?).
The Tesla K80 is a dual GPU like the Titan Z...so the V-ram is shared 24/2...so you don't have "the full 24gb" like you suppose on your first post. IMHO if you've a K80 for free you can try to use it...if you've to pay it the price I saw for used, better that card go on sleep in the shop. If you need absolutely more than 11/12 GB of V-ram there are other opportunity (P5000/P6000 and so on...)
I found 2 Tesla K40s on Ebay for ~$230 each. K80=2 K40 on one card. Granted they have "only" 2880 cores each (980 has 3200+) but hopefully I can get them both into one machine and get 5760 cores and 12GB VRAM total. I'll find out next Wednesday, I guess.
I'll also be able to tell if they work with Pascal-series GTXs.
Ok, so I've had these 2 Tesla K40s for a week and a half, and here's what I've found:
1. These particular units do not have fans on them, and they get insanely hot, even when idle.
2. According to what little info I could find on them (aside from what they do and the fact they're intended for use in servers and not desktop workstations), they require PCIe slots that are both x16 Mechanical and x16 Electrical. If your slots are marked x16(x16), that means they're the right ones.
3. TechPowerUp does not have a BIOS for them. Which means you can't directly compare it to the BIOS of a Quadro and flash it for better compatibility with non-server setups (or even to try that out).
3a. They can be seen by Device Manager, but will generate Code 12 resource errors. GPU-Z will see it, but BSODs when you try to extract the BIOS with a PNP error. NVIDIA-SMI can't see it unless the system allocates Resources.
4. I could not get rid of the Code 12 errors, even when my one system only had the Titan X (Pascal) and Tesla K40 in it. This system reads the TXP and 1080ti just fine, so I know that it's possible. "Something" is preventing the K40 from being read properly by Windows 8.1 x64.
5. My HPZ600 with dual Xeons and 48GB of RAM running Win 7 Pro x64 didn't fare any better with the K40, and it has a K4000 as primary, with the Quadro driver package loaded (previously this machine had a GTX980 for rendering and K4000 for primary, with GTX driver package loaded, and ran fine). I'm not sure where it expects to get these Resources from. Previously, on my TXP machine, I had to butcher the Device Manager stack to get my 4x GPU cluster to work, and those excess resources are still missing, so I can't say what else I need to remove to make it work. I did see somewhere that someone else had to remove 2 RAM sticks (of 4) to get one of their GPUs to work, which means the MOBO they had is set up to share bandwidth between the RAM and PCIe slots (how idiotic). I need to check my ASUS Z87-Pro's specs, as I may have that same stupidity going on. With only 32GB of RAM, I'm not looking forward to (i.e. have no interest in) cutting my system down to 16GB. Then again, if there's a 12GB VRAM limit from the GPUs, I need to restrict my scenes to that anyway. But still, 32GB gives me enough *extra* RAM for Windows to do what it needs AND actually have my full scene loaded without the pagefile annoyance.
6. For the Titan XP machine, I had to run the K40 on my Akitio Thunderbox e-GPU card, because the K40's power connectors are on the back edge, not on top like they're supposed to be, and the drive bay frame in my case won't allow for that. This device is not exactly a GPU device - it's more for sound cards and such. It's also intended for use on a Mac, which I still don't quite understand since Macs can run Windows. However, Intel makes Windows drivers for Thunderbolt, and they work well enough that the box does work like I need it to with the 1080ti, which apparently doesn't require x16(x16) slots, at least for iRay.
7. Getting a straight answer from anyone for the question "How can I get a K40 to run on an HPZ600 with a Quadro K4000 as primary running Win7x64 Pro" is literally impossible, because (actual responses I've found at various places):
a. Teslas are not for gaming (as if that were the only thing anyone wanted to do with them)
b. They have no video output (as if this was not abundantly clear from the on-set)
c. They are intended to be used in servers (but not hard-coded or physically designed in such a manner as to make them restricted to only being able to fit in server-focused motherboards).
d. Why would anyone want to use what amounts to a GTX780ti with 12GB VRAM which is now going cheap at $230 each when they could spend 3x to get a newer 11GB Pascal or 8GB RTX unit. Who saves money these days? Pfft. (Of course, looking at Ebay right now, there's a couple of 12GB K6000s with BINs under $500 each - but then only 2880 cores. Two K40s is 5760 cores for the same price. The economic sense practically punches you in the face)
e. The HPZ600's 300w PSU cannot power a K40 and K4000 at max load, even though running the 2nd GPU with an external PSU is working just fine, but who does that? Why would you want such a noisy system? Ewww eww gross!
f. It can be done fairly easily if you're running Ubuntu, Linux, or any other "DIY" OS that won't run any of the programs you've ever spent any money on, which means all that Daz content you bought is useless but you can make your own identical stuff in Blender, because obviously you know how to do that, right? RIGHT?
So basically, the lack of reliable info is astounding. I really don't care about these nitpicky details I've seen; it's a CUDA-enabled device that can be accessed just fine over a network for rendering 3D images. I can build a Windows 2008r2 server, put them in it, go through all the rigamarole of setting up network rendering, and they will dedicate to the task and spit me out a rendered image. I'm looking to cut out the middle-man and have it do the exact same thing in the same machine as the actual scene to be rendered, and no one on the internet has the details on (a) how to get it done, or (b) why - exactly - it cannot be done outside of Windows Server OS.
So, unless "the secret" drops into my Inbox, these two are going back to Ebay while I can still get what I paid.