How to test if tensor cores are working? (CMP 100-210)


I bought a used ex mining CMP 100-210 card, it works in CUDA apps (stable diffusion, Fooocus, LM studio). I use the tesla v100 driver, it works in tcc and wddm modes.
I know most CMP cards probably have their tensor core disabled.
But I don’t know how to test if the tensor cores work or not?
Under windows if possible.

And is there a technical documentation for this card? (for the V100 PCB), the card is limited to pcie 1x mode, there was a missing 2x2 SMD-s on the pcie lanes that i repaired so i think it is now just a bios or board strap limitation. So i looking for any info about the board strap smd-s. Worst case i set the straps to the V100 configuration but it is better to know what settings are there by default.

I doubt you will get information about bypassing hardware restrictions here.

There is cudaTensorCoreGemm sample code on github: cuda-samples/Samples/3_CUDA_Features/cudaTensorCoreGemm at master · NVIDIA/cuda-samples · GitHub

I just want to use this card, because it is now just ewaste.

cublas sample

Does your country or state not provide a recycling system for electronics? As far as I know, these “mining” GPUs support DX12 so you should be able to use them for gaming.

It can run games with the stock unmodded driver, but the pcie limits the performance. (it works with the quadro, TitanV and Tesla V drivers but it needs to be manualy installed)

some syntetic bench works, but real games like cyberpunk it is limited to 10-12fps max (all low and high settings)

If you profile the cuda sample with Nsight Compute, it will give you an indication of Tensor usage.

Plug the GPU into a PCIe gen3 x16 slot, rather than a x1 slot (as indicated by the GPU-Z output shown above). Depending on application, the bandwidth of PCIe gen3 x16 can already act as a performance limiter, let alone cutting that by a factor of 16.

The OP indicated that these cards (mining models), have only 1x interfaces.

According to the information I have available to me, these GPUs do have a PCIe 3.0 x16 interface, but miners chose to plug them into x1 slots to maximize the number of mining cards per mining rig. That information may be incorrect; I have no hands-on experience with these GPUs.

A GPU limited to a PCIe gen 3 x1 interface is pretty much useless except for very niche applications like crypto-currency mining. I recommend recycling these electronics unless your time has zero cost.

The card is not limited to mining, i had a few mining cards. A few them are good for low budget gaming. (P106 (GTX1060 and 40HX ( ~RTX2070) can work in 16x mode)
Making an RTX 2070 from a Mining CMP 40HX for $100

I use this card for a few cuda acceleration. Faster pcie mode can accelerate the model uploadig, and gaming.

Agree, its’ a waste of a good card. This post indicates they’re crippled at 1x.

Crypto-currency, including the creation of any kind of “mining hardware” for it, is anathema to me: A colossal waste of resources all around.

It is no surprise though that hardware vendors jumped onto the bandwagon created by the greed of crypto-currency miners. Where there is demand, there will be suppliers desiring to profit from it. During the California gold rush (1849) many of the entities that profited the most were equipment suppliers. That enabled Leland Stanford to set up Stanford University, as far as I know.

I don’t care about the miners, but i like the cheap used cards. It’s sad that the cards are very limited in many ways. This is a great GPU, which is still very capable today. HBM2 16Gb vram, over 1000Gb/s bandwidth… but I don’t have all the knowledge to restore it or convert it to V100. (yet)
So what i learned from google searching.
Board straps on the back of the card, i can set it to the V100 configuration.
Bios, i can reflash it with external flasher (ch314a) to the v100 bios. Hacked flashers not working (omgvflash and nvflashk)
Driver, there was a source leak back in 2022) but i think the driver just reads config from the GPU. (and that is not legal to mod it from that source)
Still there is the evil Falcon protection in the GPU…that can block the gpu
It is an EOL card (GPU is from 2017) so I don’t think that nvidia cares.

It is not just about the card, it is a fun project for me. I can buy a used 2080 22Gb modded card for my AI projects that has the same performance, but I don’t want to.

I will try to set the 0R SMD-s above the pcie caps like the tesla V100. I hope that controls the pcie settings.
TitanV/V100/GV100 settings

My card:

Those smd-s are controlling the board id. that is not needed for my mod (i can force flash with different board id, i need to mod the gpu id somehow to get working as a v100.