Nvidia CMP 100HX (CMP100-210) Tensor cores working?

[masterdeejay]

On the Nvidia CMP 100HX or the other name CMP100-210 12Gb are the tensor cores disabled or not?
I want to use it as an accelerator for LLM (mistral 7b medium uncensored), and it is much more usable if those cores are not fused out. It is only 12Gb but with hbm2.
Have anyone any info about theese cards?
I ordered one from ebay, it is on the way, i hope it will be faster than a P102 (~1080). Nvidia CMP 100-210 12GB Mining GPU 44 MH KAWPOW RVN | Fast Ship, US Seller! | eBay
I have used mining cards before and they worked fine for cuda, stable diffusion etc. (P102, P104, P106) And even for light gaming. But i have never used a CMP series.
This CMP 100-210 card uses the Tesla V100 PCB. (the CMP100-200 uses the Titan V PCB)
GPU is the volta GV100 i think.
For cooling i have a few ideas (i used an M40 card, it is the same airflow layout).

 

[gsrcrxsi]

I know with the Ampere HBM based mining cards, nvidia locked them down pretty hard and they were basically useless for anything besides mining.

not sure if the Volta ones like this have the same limitations.

 

[doctuspi]

Hi i was wondering if you got any update on the card or if you got it working, as found myself in similar situation, while i would find good to have tensor core enabled, i also hope for the FP64 performance to be unlocked as damn would it rip through my CUDA code. Ill also keep looking and will update here, if i figure something out.

 

[doctuspi]

So as far as the gpu is it is limited on its FP64 compute, not sure how the half precision is, but stable diffusion worked on it for me. Will need to find a way how to benchmark it though, as currently I dont have any comparative data. One thing I wondered did your card also report that it has 16 GB memory, as mine after instalation is reporting name with 12 GB but all programs see it as 16 GB

 

[masterdeejay]

My card is not arrived yet. Can you post a gpuz or hwinfo screenshot?
What driver do you use?
For compute benchmarking there is the aida64 gpgpu benchmark.
Thanks!

 

[SDletmk]

I believe there is a Stable Diffusion benchmark list you can compare your results with here: SD WebUI Benchmark Data

I am very interested in seeing how these compare with the P100 and others of a similar price point.

 

[doctuspi]

My card is not arrived yet. Can you post a gpuz or hwinfo screenshot?
What driver do you use?
For compute benchmarking there is the aida64 gpgpu benchmark.
Thanks!

Sure thing, I have wrongly named the card (it is CMP 100-210 but when adding card name to display I misslabeled it) when i modified the 551.61. driver, with NVCleanInstall

 

[doctuspi]

I believe there is a Stable Diffusion benchmark list you can compare your results with here: SD WebUI Benchmark Data

I am very interested in seeing how these compare with the P100 and others of a similar price point.

Ill install it and test it as I do have a P100 as well so can compare it to it. So far the issues might arrise with it being limited to PCie 1.0 x1 speeds.

 

[SDletmk]

Ill install it and test it as I do have a P100 as well so can compare it to it. So far the issues might arrise with it being limited to PCie 1.0 x1 speeds, but one thing is certain its FP64 compute is cut down to around 700 Gflops.

I mainly care about the FP16 performance.
That's what's necessary for Stable Diffusion if I'm remembering it correctly.
Also, 1.1x shouldn't harm that too much since all the work is done on the GPU.
Thanks for checking it out!

 

[masterdeejay]

Sure thing, I have wrongly named the card (it is CMP 100-210 but when adding card name to display I misslabeled it) when i modified the 551.61. driver, with NVCleanInstall
View attachment 35335

Looks good! Full 16Gb with nice bandwith. Default clocks a little low. No locked cores..
Pcie speeds only matters in gaming. But i only want to use it for computing, LLM and SD.

 

[doctuspi]

Here are the results for the SD WebUi Benchmark for the CMP 100-210 and P100. It seems that the tensor cores despite being present are severely limited as the P100 for me resulted in faster performance. This could also be driver related but not entirely sure as there is alot of variables.


Also here for comparison the AIDA64 gpgpu benchmark, between the P100 and CMP100-210

 

[SDletmk]

Here are the results for the SD WebUi Benchmark for the CMP 100-210 and P100. It seems that the tensor cores despite being present are severely limited as the P100 for me resulted in faster performance. This could also be driver related but not entirely sure as there is alot of variables.

Those look like about the right results for a P100, and knowing that Nvidia disabled other high-spec CMPs in a similar way, it seems about right for the CMP as well.
For the power it uses, the CMP 100-210 can't really compete with any other card for SD despite being 16GB. Nvidia really screwed these CMPs over.
Thanks for testing! It saved me from a purchase that wouldn't have done anything for me or at least some fear of missing out by not getting one before I knew it wouldn't work for my purposes.

 

[doctuspi]

I am glad that it helped. I got it to test if the FP64 performance worked on it, while i knew that likely not, it still was intriguing to get. I likely will sell it and get second P100, or Titan V as got a good find on one of them. As it stands the CMP100-210 is only good for FP32 compute, which for now not sure what it would be beneficial for atr least for me.

Those look like about the right results for a P100, and knowing that Nvidia disabled other high-spec CMPs in a similar way, it seems about right for the CMP as well.
For the power it uses, the CMP 100-210 can't really compete with any other card for SD despite being 16GB. Nvidia really screwed these CMPs over.
Thanks for testing! It saved me from a purchase that wouldn't have done anything for me or at least some fear of missing out by not getting one before I knew it wouldn't work for my purposes.

 

[masterdeejay]

My card. With a dell sff cpu cooler. Using stock tesla v100 driver manual install.




Missing pcie smd capacitor.

 

[masterdeejay]

Tensor cores are working.


Pcie lanes are repairable i think.

Little OC, using unmodded TitanV drivers manual install and msi afterburner.

 

[bayleyw]

I don't think the tensor cores are working. I get 17 its/sec on SD 1.5 on a V100-16GB without TensorRT.

 

[masterdeejay]

SD 1.8.0 RC

Dectivating unet: [TRT] v1-5-pruned-emaonly
[00:06<00:00, 2.93it/s]

Activating unet: [TRT] v1-5-pruned-emaonly
[00:02<00:00, 8.00it/s]

is there another way to test it? (on windows 11)

 

[bayleyw]

Yeah that's on par with P100 performance, which has no tensor cores.

 

[masterdeejay]

Thats bad.
I will try the modded drivers.

 

[iron-bound]

If you try Linux, may be able to modify the open drivers GitHub - tinygrad/open-gpu-kernel-modules: NVIDIA Linux open GPU with P2P support