Say no more.... What's the point of translating code that runs poorly?
AMD HIP translates 96% of CUDA code automatically. That rest 4% is your required optimization. And it translates this code not only for AMD GPUs, but ALL GPUs.
That's the price for all of these tools, performance is usually very poor. They also don't work on every project out there and good performance is usually only achieved on proof-of-concept projects. The rest is manual work that has to be done. This might be worth it when building a software base, which will serve for years to come. Doing it for projects to verify results or try some modifications is usually not worth it. I'd say maybe we keep using 10% of code from other projects and the rest ends up in the trash after we've tested and played with it. The timeframe for this is usually 3 to 6 months max and then the cycle repeats.
Have you translated any code from CUDA to ROCm OpenCL with HIP to say something like that?
The things you know are very little unfortunately. Drinking some Kool-Aid for any manufacturer doesn't really get you very far in the real world. I have no doubt AMD works for your specific area of application and that's what matters for each individual case. Keep believing everyone is wrong and you're the only one right. If you research, publish, have to colaborate with people from around the world on different projects it becomes somewhat apparent that there are solutions out there that might work in some cases, but not on others. If it's so great and works without problems, why isn't everyone using it? Why are we pretty much locked in a NVIDIA eco system to be able to do the things we need to do without adding tons of work as overhead? Unless everyone steps out of it at the same time, nothing is going to change. AMDs GPU marketshare fell from 2016 to 2017. We'll have to wait for the 2018 numbers but I doubt it'll climb.
I asked you a very specific simple question.
That might indeed be the case, yes. But the problem is, there are good reason to choose to be locked in (see my other posts). So AMD has to seriously up their game and even then, it's not something that's gonna happen in a year, or two or anytime soon. I see this among my colleagues (focused on big data), most are not even willing to consider anything else due to additional work required. We need real alternatives that work with the push of a button.
In properly optimized code situation, AMD GPUs are as fast as Nvidia's. No matter what AMD will do with their hardware(ALUs are just that - ALUs). You haven'ported properly your code, or even used AMD GPU with properly ported code if you do not know this. There is plenty of things that AMD does not have solution for, yet - lack of libraries. And those cannot be ported as easily as simple CUDA code.
I see this a lot, curious what we'll do for our clusters. I'll try to rent in the cloud for my projects in the near future, but can't obliviously speak for everyone in the world (over even my colleagues). Just need the Quadro cards in some cases. Then again, since the whole world has no clue what they're talking about, maybe everyone should be educated and learn how to do things properly. Then NVIDIA would be bankrupt next year and we wouldn't have to buy these expensive cards.
I'm curious about the price the big cloud operators pay for Volta cards.... Maybe Nvidia sees cloud GPGPU clusters as the future, and is heavily discounting the cards for AWS/Azure.
I briefly mentioned this, if you're starting new or your code base will remain for years and you don't plan to switch, pick whatever suits you most. If you don't need to port, if you don't need to scale to huge clusters and if you don't need support from the vendor, it doesn't really matter, assuming no platform is discontinued at some point.
I don't disagree - but this comment seems to assume that one is creating an independent standalone application.
Which is why you ultimately are wasting money. Because once you port your CUDA code using HIP to any other paltform, and you are good to go with it - everywhere. AI and ML are very well supported on AMD platform, and Vega 64 is offering 95% of performance of GV100 chip, but at the same time costs 1/6th the price. Buy 30 GPUs at once and you get the difference(900 000$ vs 150 000$).
It's what I assumed, yes. You're absolutely right about what you say. I'm active in the same field (my original background is in graphics and image processing with 10+ years focus in medical image processing though) and it seems like we're using the same tools. I agree with what you say, it just saves so much time.
