Nvidia CMP 100-210 or 100HX (GV100 GPU)

[masterdeejay]

I bought this card from ebay. So this topic is about the CMP 100-200/210
My card was cheap (160usd)
First this is a dedicated mining card so expect a few or a lot limiations compared to the retail gv100 cards,but still usable.
For drivers we can use the newest TitanV or the Tesla/Grid drivers (switchable wddm/tcc mode)
It uses the 8 pin EPS but i got an eps to 8pin pcie adapter.

The spec:

I have tested a few things:
Fooocus and stable diffusion works well aprox 2080ti/3070 performance, the tensor cores are working but lower clocks than the tesla v100
Kobold cpp works in cuda mode.

Pcie is limited to 1.1 1x (dont know that is a hw or sw limitation)
So gaming performance is very limited and the modell upload is slow.
Double precision compute is limited.

TDP set to 250w but it is maxed out at 190w. It runs at 60c max.

Overclock is possible with msi afterburner, but it is not stable at the v100 32gb clocks.

For cooling i use a cheap blower fan from a dell sff pc.

Pcie 1x mode reason is the missing smd.

 

[doctuspi]

Hi, it is nice to see another post made on these card, as there is not much information on them. I was curious how did you got SD working on tensor cores, as when I was attempting to run it (my results here in the discussion on ServetheHome website), it only gave me performance similar to P100 that I also had. For me the double precision was a bummer, as that was my biggest interest, but if the tensor cores are working and I just had a bad setup then it still looks like a great card.

With the missing SMD, maybe if they get soldered back on it might enable higher PCIE than 1x, I can look into it and buy them to test it out. All of them with initial multimeter measurements seem like 50 MOhm SMD. I will remeasure it later with my second multimeter to check. If this card can be modded with solderig to gain additional pcie lines than it might be decent for even gaming as the FP32 performance is not cut.

With the drivers, have you performed any modifications to them, or followed the same modification procedure by adding the card to the drivers.

Lastly I do hope that the bios flash program gets patched for these, as if flashing V100 would unlock further abilities of this card, and soldering the SMDs would exapnd the PCIe lanes, then it would turn from decent card into an amazing one, especially for the money (the prices though rose on ebay as they are now around 310$).

Also not sure if it will be useful to anyone but here are some bare board pictures of the CMP100-210.

 

[masterdeejay]

SD tensorRT extension installed. Enabled, disabled and again enabled.
3 it/s disabled 7-8 it/s when enabled

Loading TensorRT engine: T:\AI\stable-diffusion-webui\models\Unet-trt\v1-5-pruned-emaonly_d7049739_cc70_sample=2x4x64x64-timesteps=2-encoder_hidden_states=2x77x768.trt

Loaded Profile: 0
sample = [(2, 4, 64, 64), (2, 4, 64, 64), (2, 4, 64, 64)]
timesteps = [(2,), (2,), (2,)]
encoder_hidden_states = [(2, 77, 768), (2, 77, 768), (2, 77, 768)]
latent = [(2, 4, 64, 64), (2, 4, 64, 64), (2, 4, 64, 64)]

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 7.84it/s]
Total progress: 30it [07:30, 15.00s/it]
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 7.71it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 6.71it/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 7.30it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 6.88it/s]
Dectivating unet: [TRT] v1-5-pruned-emaonly████████████████████████████████████████████| 20/20 [00:02<00:00, 8.02it/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:06<00:00, 2.93it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:06<00:00, 2.94it/s]
Activating unet: [TRT] v1-5-pruned-emaonly█████████████████████████████████████████████| 20/20 [00:06<00:00, 3.01it/s]
Loading TensorRT engine: T:\AI\stable-diffusion-webui\models\Unet-trt\v1-5-pruned-emaonly_d7049739_cc70_sample=2x4x64x64-timesteps=2-encoder_hidden_states=2x77x768.trt

Loaded Profile: 0
sample = [(2, 4, 64, 64), (2, 4, 64, 64), (2, 4, 64, 64)]
timesteps = [(2,), (2,), (2,)]
encoder_hidden_states = [(2, 77, 768), (2, 77, 768), (2, 77, 768)]
latent = [(2, 4, 64, 64), (2, 4, 64, 64), (2, 4, 64, 64)]

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 7.88it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 6.85it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 8.00it/s]

 

[agent_x007]

Try to find [CUDA : Force P2 state] option under NVidia Profile Inspector, and set it to [OFF].
This should help with performance under some of compute workloads.

Does V/F curve control work ?
If they do, try setting ~1620MHz at 0.85V and see if it helps on power throttle side.

PS. x1 3.0 WILL limit performance A LOT, regardless of what you plan to do with it.
Good thing those caps are standard ceramic ones (no polarity), just "borrow" few of them from another (or dead) PCI-e 3.0 card (that has similar sized ones).

 

[masterdeejay]

 

[masterdeejay]

The missing SMD-s on the pcie line is 200-220 nF capacitors. I will solder them back.

For driver i use the Quadro GV100
Superposition only runs in windowed mode and lags a little at the begining but i think the scores a bugged.
Temperatures not reported correctly.

 

[agent_x007]

Here's Cuda-z from my Titan V (PCI-e 2.0 x16 limited :/) :

^Highest F/V point : 1665MHz @ 0.8V

Superposition with F/V curve, no VRAM OC (in fullscreen) :


1080p extreme (window mode) :

^Yes, 1080 extreme is heavier GPU wise than 4k optimised.

 

[Solaris17]

The missing SMD-s on the pcie line is 200-220 nF capacitors. I will solder them back.

Excited to know how that goes. Keep us in the loop!

 

[masterdeejay]

New 2x2 220nF 16V 0402 X7R SMD caps.
I still get 1x mode. So it is a bios limit.
I searching for board straps or some documentation.

 

[masterdeejay]

I tried a few things: modding straps, modding bios... the card is locked with the falcon protection in the gpu... so it is a good GPU but crippled with gpu fw lock...
I gave up the project.

I sold the card for 270 usd (so ~100 usd profit), and i bought a used RTX 3080 FE...