Hacker News Reader

This is what many people outside the AI world don’t seem to understand. Nvidia has a stranglehold in the form of CUDA and Cudnn. There isn’t any open source equivalent to Cudnn. AMD is trying to push OpenCl in this direction but it will be a long time before DL libraries start migrating to OpenCl. Like tomorrow by miracle if al alternative GPU which is as good as the 1080ti popped up, it would be useless in the AI market.

View on HN

No it won't. Especially if the price is competitive. Say for the price of 1 1080ti, if i can buy 1.5 units of comparable performance graphics card, i'll surely buy it. There are already resources being spent on OpenCI based ML/DL platforms (https://github.com/plaidml/plaidml). The architectures keep getting bigger and training time keeps getting longer. I think you underestimate this factor. I need as much gpu computing power as i can buy within the budget.

True, I would love for some alternatives such as plaidml. However, I can't seem to fathom that Plaidml will be a worthy alternative to lets say Tensorflow or Pytorch or Caffe. I hope I am proven wrong.

I think support for OpenCI will eventually come to other frameworks. But the main problem is the fact that AMD is still far from the NVidia in terms of performance. Vega couldn't reach the performance of 1080Ti, and with Volta next year, the gap is gonna drastically increase. If only AMD can fill the gap, the support, i'm sure, will soon come after that.

Reply on news.ycombinator.com