Are NVIDIA CMP GPUs any Good for Deep Learning?

Soon, NVIDIA will release the new CMP line of GPUs specifically designed for crypto-miners. Naturally I wondered what use these GPUs might have for deep learning practitioners. In this video I take a look at the specs, introduce the line, and discuss why there might not be much overlap with the needs of deep learning.

NVIDIA CMP GPUs:

** Follow Me on Social Media!

[EvgenMo]

It's been a couple of years now, has anyone had any experience with these CMPs??

[Combatwhombat]

CMP-170HX vs A6000
Yes, they actually should be good for DL. And are Ampere GA100's.

Shading Units 4480 vs 10752

TMUs 280 vs 336

ROPs 128 vs 112

SM Count 70 vs 84 (streaming multiprocessors contain 32 tensor cores)

Tensor Cores 280 vs 336

L1 Cache 192 KB (per SM) vs 128 KB

L2 Cache 8 MB vs 6MB

When running an A6000 for DL/ML the shading units would be dark silicon, representing wasted expense on parts of the chip that are unutilized or underutilized. The CMP simply trades these for some more tensors and a massive 4Kb memory bus compared to 384 bit on the A6000. You see, as you mentioned that for mining the bandwidth between storage, the CPU, and ram is not significant in performance (though 286 is a big stretch). Within the cards, the memory speed and bus are a greater limiting factor in ethereum mining than the computational prowess of the CPU, but this also implies that the memory bus speed should benefit in computational throughput in AI/ML/DL.

Provided that NVidia doesn't lock the drivers of the CMP cards from compatibility, they should work but not as a dual-purpose gaming rig. Assuming that this is a relatively straightforward adaptation of the same architecture, there isn't much incentive to lock access to CMP for only mining; these should be just about pinnacle deep learning workstation accelerators for now. But, it does appear that they cannot be used to output a display, which makes it a secondary card.

[kektus]

Loved to hear your views on the topic Prof. Heaton! Thank you for the video!

[chrisminnoy]

I agree with your analysis. If the CMP line of hardware would favor both the miners AND the deeplearning engineers, there would at least be a second hand market. At this moment the CMP hardware has no use when the difficulty of the mining algorithm goes up.

[jedcheng]

It is going to be a nice one after a crypto market crash. I have heard people buying second-hand p106 (the mining version of GTX 1060) after the crash in 2017 for GPU servers

[Christopher-today]

Hey Jeff, can you do a video talking about these 2 topics...

1. On the A6000 with its 48GB memory can you give examples of the biggest models (image & text) on the signal card?

2. Can you explain how you determine the approximate time required to train a model on a given GPU plus the other inputs that effect the training speed?

Thanks so much,

Christopher

[Brillibits]

I agree that a RTX A6000 would be a poor choice for mining. However, I am curious, could you please tell me what mh rate you get on it? I can't find this easily on the internet.

With the way things are going in the GPU market, A6000s may be used for mining soon. My RTX 3090 FE has a resale value of over 3k atm due to miners. I am half tempted to sell it and get an A6000.

[BlueprintProgrammer]

So did anyone tried using these for deep learning? What speeds are you getting and is there any problems?