MIOpen: AMD's Machine Intelligence Library | Hacker News Reader

MIOpen: AMD's Machine Intelligence Library

github.com

131 pointsjonbaer22 comments

I continue to buy AMD cards for gaming because of their relatively greater commitment to openness.

I buy NVidia because while closed, their drivers work better on open operating systems.

I think the real story is the HIP library talked about. Automatically translates your CUDA core into a more C++-like language with the same features.

I thought HIP was an incomplete translation. Their documentation said something about HIP doing 99.6% of the translation and one developer doing the rest in a week. Unless they've improved this, i'm not sure how much value HIP brings.

edit: I just keep wondering why AMD doesn't put money into developing an OpenCL version of CUDA/CuDNN

Something that is interesting is that it's not only a translation system but also it's own API which looks much nicer. It also supports Cuda and OpenCl it looks like.

MIOpen runs both OpenCL and HIP so aka in your words is the "OpenCL version on cuDNN".

Oh wow, this is way too cool. Just 3-4 days back I was thinking I should invest in NVDA as they pretty much have a stranglehold on AI because of CUDNN and CUDA. I couldn't muster the courage to buy NVDA considering how much the stock has grown in the last 12 months. Now there seems to be an alternative. Perhaps I should start buying AMD. The repo has 21 contributors and I just checked them randomly and most of them are from AMD. It'll be interesting to watch the contributions grow from outside AMD

> considering how much the stock has grown in the last 12 months.

You might have the same reservations about buying AMD.

AMD might be a better investment because of the stock price and their resurgence.

They aren't close to threatening NVIDIAs hold over GPU accelerated ML/Compute market just yet.

Don't buy individual stocks. Less than one percent of people have the ability to choose individual stocks that perform better than mutual funds.

Do you have any source on this? That statistic does not make any sense.

Given that literally copy/pasting the sentence "less than one percent of people have the ability to choose individual stocks that perform better than mutual funds" into Google found a reference, it probably took me less time to find this than it took you to scoff in disbelief and ask that question :(.

http://www.marketwatch.com/story/almost-no-one-can-beat-the-...

(Note that this article mentions multiple slightly different studies, so if you are the kind of person who is going to glance at one paragraph and then scoff in annoyance that it doesn't sound like quite the same statistic, I recommend you keep reading. You can start to quibble at the end ;P.)

I've been bought into AMD for a long time. I've said this so many times I'm starting to sound like a shill for AMD, but I respect their new CEO so much. I really think she's taking the company in the right places, because she's an engineering-minded CEO, not a business-minded one and I think she's exactly what AMD needs right now to compete.

Will this make it possible (in the long run) to train tensor flow models on AMD acceleration cards?

It's already possible with some unofficial libraries.

Ok, allow me to me rephrase GP: will this make it possible (in the long run) to train tensor flow models on AMD acceleration cards in a supported, easy-to-use and performant way, comparable to what NVidia is offering?

I think that's the goal.

So this looks like the AMD alternative to the Intel DAAL libraries?

Pretty cool that they're offering a HIP (CUDA) version too

It looks like AMD's version of CUDNN. DAAL seems to be more big data than machine learning.

If they manage to add support for all the major frameworks it would instantly become a serious option.

Right. CuDNN -- AMD MI -- Intel MKL-DNN.

Reply on news.ycombinator.com