Anyone have experience with 2 x Tesla M40 24GB?
necrophyte69
Posts: 17
I purchased two of these thinking cool plug them in and go. Plugged in the first one to make sure.. After some googling, found I needed to enable 4G Decoding ok cool No problem. One down. Works great renders awesome large scene. Great. Install the second one and... No POST.
Ok pull them both out, try the usual. So the second one works no problem. No combination of moving 1st, 2nd, or 3rd slot work. Just can't get 2 to work. tried disabling everything in the BIOS, setting everything to GEN 1. I'm just pretty sure at this point its not possible in a Z370 to have two work.
So here is the setup in case someone has questions
Asus Prime Z370-A w/ i5 8600K
EVGA 1000W PSU
I have everything else out for troubleshooting purposes. Any ideas?
Willing to purchase a new MB and CPU if I have too. Just need to know what for sure and no one can seem to help and internet is very scarce on information about these and what will or won't work.
Any help is appreciated,
Thank you.
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
Hmm doesn't sound promising... do you have any other PCI cards installed that you can possibly pull? I assume not.. but other than that, I think your only option is to find another motherboard to try.
Tesla's are massive resource hogs. You're going to need a server grade Mobo to get two of them working in the same system. Nvidia probably has a list of MoBo's that support them somewhere.
IOW if you want to stay on desktop hardware you won't get both running.
Nvidia does have a list, but they are all rather old MB's that I can't see to find anywhere. Does anyone know if maybe an X299 with a 7800X could do it? I know they have some 28 PCIE lanes. While the 8600K has only 16. Was thinking maybe thats whats holding it up. I have a spare X299 motherboard sitting around, but no CPU atm. All the CPU's I see caple of handle such beasts apparently have 40+ Lanes. 7900X has 44 Lanes. I'm just not sure if thats what the issue is or not?
So I thought I'd put it here in case someone else ever wanted to know. The folks at EVGA say it will work. So I'm going to pick up the CPU soon. I'll let you guys know what I find out once I get it.
Really? I'd check with Nvidia before buying the CPU. Tesla's require some weird PCIE provisioning that I spent the better part of a day trying to figure out before I got lost in UEFI stuff I simply hadno experience with. I'm tempted to just go with number of PCIE lanes but I'd hate to see you spend the money and come up empty again.
Check what has been done here https://www.servethehome.com/nvidia-smi-issues-get-nvidia-cuda-working-grid-tesla-gpus/
I doubt anybody has Tesla multigpu experience with an i5. That's a very uncommon configuration.
The PCI Lane limitation can be the culprit but you're the only one to be able to test that. Make sure that the two PCIe lanes get 8x/8x when testing. That would be the first prerequisite
8x/8x at least. If you can arrange 16x/16x I'd try that as a first option. But I still have very strong doubts. That Supermicro stuff looked like the stuff I was researching. I don't think any desktop UEFI I've ever seen has those sorts of settings.
nVidia's response was to buy one of their servers, and that they don't support anything other than those.
I sent the guys at supermicro a message for a setup that would support the 2 x M40's (Was willing to buy it). But they haven't responded in over a week now.
Going back to the i5 I was able to get a 1080 TI and 1 M40 to POST. I didn't go much further that that as I know there are issues with the GTX and Tesla/Quadro drivers. When I would put the 2 x M40's in it wouldn't even POST. Just halt on VGA.
Even if I can only get one of these working it was still worth it. eBay had 5 of these up for $500 USD ea (New) so I grabbed 2. Of course once I got them I plugged them in and .. Nothing.. Lol. Go me for not doing enough research and thinking that I knew enough about video cards that why shouldn't 2 of these just "work".
Anyways, I just waiting for the parts to show up and I'll update here with what I find and if it works or not.
you could always just have two rigs one for each card I guess
Even just one should be a massive boost to your render speed. 3072 CUDA, even a couple of generations old, and 12 Gb of VRAM is pretty impressive.
I do know that Supermicro plans to get back into the desktop motherboard market sometime this year so even if you can't get them running immediately I'd keep an eye on supermicro because if anyone is likely to put out a board with the sort of UEFI settings you need it would be them.
A Tesla has no video outputs so it cannot be your only graphic card also if you keep an eye on eBay you can get Tesla's for much less than MSRP. Is it faster than a 2080ti? Almost definitely not. The M40 is 2 generations old and has 3072 CUDA. The 2080ti is the newest generation and has 4352 CUDA. The M40 does have 12Gb of VRAM which would let it render scenes too large for the 2080ti.
As to going from 6 to 8 or 11 or more Gb of VRAM yes it will make a difference. The reason your card drops to CPU is the scene doesn't in the graphics card's 6 Gb so adding more VRAM might have kept the GPU in the render. Further the higher memory cards have more CUDA which will give you faster rendering.
Yes, there are newer Tesla's with more VRAM and more CUDA I was only discussing the M40. I have no idea what that dev was talking about though. VRAM lets you fit the scene on the card which is vital to letting the GPU render at all but CUDA is what determines how fast the GPU renders, until Nvidia enables the RTX features on the RTX cards in iray. Also keep in mind that CUDA is only directly comparable in the same generation of graphics cards. Pascal to Pascal and Maxwell to Maxwell. Performance per CUDA has gone up every generation so comparing Maxwell to Turing isn't anywhere near as simple as directly comparing CUDA.
The Tesla M40 should be slightly better than a GTX 1070 in term of processing power. So, you won't have that much rendering power if you compare to Turing Cards. But for around 1000$ that's still a good deal with 24GB Vram if you need as much memory (big scenes). If you only do small/medium scenes, then I don't think it's a good choice and I'd rather go for an RTX 2070 or 2080.
No problem.I just want to make sure you know you won't get RTX performance. Beside, it's your money you'll be spending, and you know what you want to render. The choice is yours
Ok So I have an update.
I decided NOT to go with the Intel 7900X with 44 Lanes. I was able to purchase an AMD 1900X Threadripper with 64 lanes and motherboard, for a fraction of the 7900X just the CPU.
I can confirm that 2 x M40 24GB cards work with an AMD Prime X399 and an AMD Threadripper 1900X. Still have enough room for lanes to use the M.2 drive I have as well.
I grabbed a Quadro 2000 for a display.
So now I have two cards that can render a scene with 24GB of Memory
According to a chart for rendering iray on this site. An M40 is faster than a 1080 in rendering. Its not faster than a 1080 TI, but does it really matter how fast a 1080 TI is if your textures require 12GB of Vram?
I'll take 1080 Speeds with 24GB of mem anytime!
It was trickey getting drivers to work. You need to make sure you when you go to the nvidia driver website that you select the exact card you have for display. Versions will change. So make sure you got the right one. Took me a while as I assumed they were like GTX drivers and you can just say its a 10 series even if its a 9 series.. I was wrong. Then I installed the Tesla drivers using express. If you use clean it kills the quadro drivers and your stuck having to go back and install the quadro drivers again.
Rendering is faster than a 980TI, but slower than a 1080TI. I don't have a 1080 to compare it to.
Overheating is a thing. Make sure you have a fan in front of them (Front of the case). I also placed a 90mm fan (Exhaust) on the back of the case with some 3M doublesided tape. It seems to have solved the issue.
A Scene I rendered with CPU (8600k) before was just shy of 19 hours, w/dual M40s was just over 4 hrs (4hrs 12min). Single was over 8hrs (8hrs 32min).
EDIT:
They should have taken less time, but I'm dealing with heat issues, and the cards throttling. Working on sorting this out.
Another Update, I was setting up my wifes new machine. Just an i5 8500 w/H370 motherboard (Asus H370M-PLUS). I was actually able to get it to post and run with 2 of the M40''s. I've tried it in 2 Z370 Chipset boards, and a Z390 Chipset and it didn't work. Also figured out the heating issue. A 980TI Reference card shares the exact bolt on cooler. So any 980TI/TITAN X (Maxwell) cooler will work. The Corsair HG10 N980 is the best/cheapest solution I could find. Cooled down the scene only took 2.5 hours.
@Kenshaw011267 These Tesla M40's are actually 24GB of ram. Great for rendering massively large scenes!
Interesting! So until you enabled 4G decoding, even a single M40 would not post?
Thanks!
how are you cooling these cards?
There is a new discussion about these cards that includes cooling: https://www.daz3d.com/forums/discussion/402991/successfully-egpu-d-a-tesla-m40
Trying to avoid replacing the heat sink, although (from notes above) this is also an option.