Can we lobby to keep K80 support (sm_37) in CUDA 12?

Can we lobby somewhere at NVidia to not drop them in CUDA 12? They’re such wonderful cards, I have six of them, and it seems they’re still widely used by many corporate clients and data centers, incl. the P2 instances in AWS. It would be a pity to drop the K80 (and K40 presumably too, that’s sm_35). If anyone involved in the decision-making on this, here is my +1 vote to keep it!

(1) Even if all remaining sm_3x support will be removed in CUDA 12, the hardware would remain perfectly usable with current software. For example, the system on which I am writing this has an sm_30 GPU with a driver from 2019 and CUDA 9.2. It still works perfectly fine running CUDA-accelerated code.

(2) There are a bunch of supercomputers that use K80, e.g. SDSC’s Comet machine which is scheduled to operate at least through March 2021. As I cannot imagine NVIDIA pulling out the rug from underneath large existing installations, I would expect [speculation!] K80 to remain a supported platform for at least another half year.

Given that GPU architectures are still evolving at a pretty good pace, at some point it makes little sense to offer the latest software with support for “ancient” hardware, where “ancient” currently resides near the six-year mark. A mark that K80 will reach in a few weeks.

Thank you for the info.

Yes, I understand the hardware will continue to function, the only thing that will change is the ability to compile on a newer CUDA version. And as it’s possible to have different CUDA versions on the same machine, it should be possible to keep a CUDA 11 intact for the K80s and install a new CUDA 12 separately on the same machine for Ampere. I could then use that to compile for a GTX 3090 while leaving all the K80 stuff intact.

And I clearly see the advantages of tensor cores, so I’m also interested in modern cards and modern CUDA, but not all CUDA applications need tensor cores, and the K80s simply are very good workhorses still today. (and can’t beat the price)

Circumstances will differ. I live in California where electricity is expensive compared to the rest of the country. Since I run (some of) my GPUs pretty much 24/7 op-ex tends to be more important than cap-ex when it comes to hardware. I very much appreciate NVIDIA’s advances in power efficiency in their more recent GPUs and wouldn’t want to run a bunch of K80s at this time, no matter what the purchase price.

You can definitely have multiple CUDA versions installed on a system. I think I have four on this machine. Using everything from the command line makes it easier, I suppose. Not sure how multiple version co-exist in an IDE.

“I live in California where electricity is expensive compared to the rest of the country”. It’s hard for a Trumpist not to take a jab at CA politicians at this point, but I’m sure politics is not allowed here :)

To save electricity, perhaps you could use liquid cooling and send the water through the baseboard heating, especially as you’re running tasks 24/7, hence generating a lot of heat. I find it crazy that with the GPUs we pay money for the cooling (fans, or pumps for the radiator), yet in other parts of the house we pay for heating – would be cool if we could simply redirect the unwanted heat from the GPU in liquid form so that we can recapture it in the baseboard heating for the house. Oh wait, in CA you don’t need heat, in fact, you have so much of it that your forests are burning. Here in NJ we sometimes need heat, and GPUs produce just that!

“Not sure how multiple version co-exist in an IDE.”. Should I perhaps consider containers? There must be containers for the different CUDA versions, perhaps even for the various distros (Fedora in my case). That way I could use different versions of NSight Eclilpse Edition for the different CUDA versions.

Not much home heating is needed in this part of California, correct. I am sure if you talk to the people up in the mountains of California they would have a different story to tell. My GPUs give off enough heat that I did shut down several during the multiple heat waves that hit the Bay Area this year.

Speaking as an engineer, using GPUs as space heaters in a cold(er) climate seems needlessly inefficient. It is much more efficient to burn natural gas directly than first converting it into electrical energy. In any event I am happy that NVIDIA appears to be relentlessly pursuing higher power efficiency with their GPUs. Possibly just a side-effect of next generation exa-scale supercomputers requiring this to operate within a reasonable power budget, but I take what I can get.