Date: Tue, 21 Jan 2014 22:36:04 +0500 From: Muhammad Junaid Muzammil <mjunaidmuzammil@...il.com> To: "john-dev@...ts.openwall.com" <john-dev@...ts.openwall.com> Subject: Re: CUDA formats Well it would be an interesting debate I think :). OpenCL is certainly better in terms of its openness and multi vendor support. There isn't any doubt in that. And that makes it the best atleast for open source community. In terms of performance, I personally think that both are equally well. What gives CUDA edge is its libraries. There are a lot more scientific computing being done in CUDA as compared to OpenCL (at least what I have observed). One such example is that over gpgpu.org, CUDA resources are atleast three times than that of OpenCL (322:101) in actual. Regarding the Nvidia vs AMD comparison, 2 of the world's most powerful supercomputers consists of Nvidia devices where as there isn't any AMD machine in it (http://www.top500.org/lists/2013/11/). With Nvidia investing a lot in academia with its CUDA programs, atleast I don't see an end to it. But who knows, may be my predictions prove wrong. I, personally would say to keep the CUDA support open. The priority for its development can be set as low but don't tune it off. Junaid On Tuesday, January 21, 2014, magnum <john.magnum@...hmail.com> wrote: > On 2014-01-21 05:41, Muhammad Junaid Muzammil wrote: > >> Exploring the repository, I found that there are more formats supported >> for >> OpenCL as compared to CUDA. I think the additional formats should also be >> written for CUDA. What do you say? >> > > We should rather drop the CUDA support altogether because all CUDA-capable > GPUs are also capable of OpenCL. You should actually never build -cuda but > build -gpu instead. That way you'll have all formats. OTOH as all CUDA > formats are dupes of OpenCL ones you could just as well build -opencl and > have everything! Most of our OpenCL formats are also capable of > auto-tuning, while none of our CUDA formats are. > > On top of that, no nvidia card comes anywhere near AMD in terms of > performance, at least when it comes to password cracking. So I'm not very > motivated to develop for CUDA at all. It's a dead end, OpenCL is the future. > > To somewhat disprove my point you could try to complete the wpapsk-cuda > format: Add the missing GPU-side functionality that wpapsk-opencl has > (should be trivial to port). Then add device auto-tuning. Finally try to > optimize it and see if it can be significantly faster than OpenCL on a > given nvidia device. That would be an interesting project. > > magnum > > Content of type "text/html" skipped
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.