Are Tesla K80s still worth it or How much longer will Eevee support Kepler?

Hi there, I am not going to get too far into the nitty-gritty of it but I want to know if the ultra-cheap K80s are still worth it (for eevee rendering). My main issue is the kludgetastic compositor node setup (which I am not allowed to share so I cannot find out how to optimize it) cause the post processing to take 2-3 times longer than the actual render with a 1070 (which is close to running out of VRAM) which means a 5000 frame, 2 day job takes 4-6 days instead. I want to get three K80s, BIOS flash them, and batch run 6 instances of Blender to fully saturate my OCed 5960x since it only seems to use a little bit of its processing power and not all 16 threads.

My main concern is its age. Render engines are dropping support for older architectures (I believe Redshift dropped Kepler support not too long ago) and who knows how much longer it will continue to receive driver updates on Linux (I am on PopOS which makes Teslas pretty much plug and play from my understanding).

That is why I want people’s thoughts on them. I know their time is extremely limited but 12GB of VRAM and over 25 TFLOPS (especially when BIOS flashed to unlock performance) for less than 600 CAD is an extreme value proposition, especially when you already have all the tools on hand to cool them and need to buy more RAM to handle all of Blender instances… It would just suck if Blender dropped support for Kepler within the next year and I cannot find any deets on this matter.

Thanks for any input!

Eevee is using OpenGL. It is based on OpenGL version, not GPU architecture.

According to Blender OpenData:

NVIDIA Tesla K40m 318.17

By contrast, a 1660 Super scores an 850, and a 3070 scores 3300. Given that you can get a 3070 also for 600 dollars, it seems like the better option

Sorry it took so long to respond. Everyone in my household now has Covid and even though I am testing negative I still feel a crappy despite getting both boosters, and now I have to worry about Myocarditis. Fun.

Anyways…

Some how, I already knew about that and yet completely forgot. It supports the latest versions of OpenGL but only Vulkan 1.1 and Eevee is said to move to Vulkan in the near future. The M40 supports the latest version of Vulkan though.

Thanks for pointing out OpenGL otherwise I would have probably not realised the latter. I lose a considerable amount of overall horsepower by going with the M40 but I have double the VRAM per chip and at least a couple more years of linux driver support.

I am limited by CPU threads (and VRAM), not GPU performance. I have over a dozen View Layers and a convoluted compositor node setup, that I am not allowed to share, which causes a 30 second overall render on a single 1070 to jump to 90-120 seconds. Times that by 2000-5000 frames.

Sometimes I feel like I am working for Tommy Wiseau.

Outside of the initial initialization phase I only see a couple of threads used at a time during this mess.

I have a friend who is in on the project, with an 5950x and a 3090, and it is a similar story in performance. It is obviously faster but still limited by only a few threads a time being utilized. I tried running 2 instances of Blender on the single 3090 and saw nearly a doubling of render speeds but Blender kept inexplicably crashing when doing that. Maybe it was the VRAM overheating but I have not had a chance to test it again after swapping out his thermal-pads.

Ultimately, that is what I want to achieve without breaking the bank, being able to run multiple instances of Blender when rendering. I already have a beefy power supply from an older build and all of the hardware to cool three Tesla’s without blowing out my eardrums. I just needed to know if it would soon stop working with Eevee due to external reasons which Hikmet’s post made me realize that I boneheadedly already knew the answer to.

Get well soon.

Kepler supports Vulkan 1.3 version and Eevee architecture does not look that will change anytime very soon.

Thanks and Gotcha.