Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by zdw on Sunday December 24, 2017 @10:56PM
(#16002559)
AMD's ROCm suite has a tool for converting CUDA to C++ code: https://github.com/ROCm-Developer-Tools/HIP
How well it works is likely codebase and application dependent.
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by mrmondo on Monday December 25, 2017 @01:34AM
(#16003020)
I’ve been using OpenCL for some neutral net image optimisation recently, I’ve actually found it very efficient (citation obviously needed), I personally welcome the competition of AMD and it does seem like Nvidia is digging its own grave (not that I think it’ll die).
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by wyldfire on Monday December 25, 2017 @12:46PM
(#16004971)
I don't think it's that hard, but it depends maybe a bit on how long your codebase has evolved through CUDA.
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by frozenport on Monday December 25, 2017 @12:07AM
(#16002750)
Here was my experience. Everything in ocl ended up slower, with 30%+ longer (walltimes) on nvidia hardware. Also some problems with floating point precision. Real problem was there was no business case to switch because we could always make customers eat the cost of taking nvidia gpus.
The added code complexity was not worth it - we originally did it as NRE for a big client who wanted OEM capabilities to our IP.
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by katastic on Monday December 25, 2017 @12:14AM
(#16002777)
OpenCL is slower than CUDA, but it's not the end of the world.
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by mangix on Monday December 25, 2017 @01:01AM
(#16002913)
got a reference?
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by unfamiliar on Monday December 25, 2017 @02:43AM
(#16003183)
That is simply not true.
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by biafra on Monday December 25, 2017 @02:57AM
(#16003228)
Please be more specific. What is not true? That it is slower or that this is not the end of the world?
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by katastic on Monday December 25, 2017 @08:30PM
(#16006542)
Wow, people are butthurt. It's a wildly known fact in the industry. A simple google search for benchmarks or a PDF like this would have satisfied them. But they didn't want the truth they wanted to be butthurt.
https://arxiv.org/ftp/arxiv/papers/1005/1005.2581.pdf
In basically every chart and graph, OpenCL is slower.
OpenCL has a higher level of abstraction, so it's got a higher penalty for it. But, you get cross-platform support. nVidia doesn't optimize for OpenCL either because they don't want you to use it over their own competing framework.
But 'dats cool, keep downvoting, guys. Sources aren't real if you can downvote 'em enough. God, I love Hacker News.
Re:NVIDIA GeForce driver deployment in datacenters is forbid...
by grzm on Monday December 25, 2017 @08:41PM
(#16006566)
Would you kindly raise the level of discourse here? You have useful information, so thanks for sharing it. But please leave the swipes and complaints about downvotes and HN members out of it. It worsens discussion and is against the guidelines. You’ll likely receive downvoted for that, more than anything else.
But AMD doesn’t run CUDA? Is it very easy to switch a stack relying on CUDA to OpenCL? I don’t think so.