Ask HN: Why did those ideas to pool cryto mines for deep learning fail?
With cryto currency price hovering below profitable lines for most crypto mines, it seems to make so much business sense to pool these GPU resources for deep learning.
There have been quite a few attempts at it, such as https://aicrypto.ai/, https://vast.ai/, and https://www.reddit.com/r/gpumining/comments/86ofw2/rent_out_your_gpu_compute_to_ai_researchers_and/. However none of them seems to have gone anywhere.
Why? Is it profitable for a miner to rent out time on their mining systems? The reddit thread says "double what he was making by mining Ethereum, and in return he would let me train my neural networks on his computer at significantly less than what I would have had to pay AWS." That is pretty significant but I'm kinda surprised it that much less than AWS, does AWS have a really high markup? AWS has to cover taxes, vacations, datacenter construction, monitoring, duplication with enterprise level components, pay rate for electricity is higher for commercial companies and so on. Probably because our datasets are just as valuable as the neural network topology and weights, and as the founder of several companies that use deep learning theres no way I'd let our datasets go out to machines we don't control Who is doing deep learning at scale? Google, fb and nvidia? Do they really care about miners’ spare GPUs. What about the risk of leaking training data to public? Is model parallelization solved yet? Especially Without high speed connectivity?