Comments on: Reports of CUDA’s death exaggerated? http://www.pcpro.co.uk/blogs/2009/09/30/reports-of-cuda%e2%80%99s-death-exaggerated/ Blogging in the real world Sun, 12 Feb 2012 15:02:20 +0000 http://wordpress.org/?v=2.8.4 hourly 1 By: Patrick Degenaar http://www.pcpro.co.uk/blogs/2009/09/30/reports-of-cuda%e2%80%99s-death-exaggerated/comment-page-1/#comment-122284 Patrick Degenaar Mon, 07 Dec 2009 12:55:31 +0000 http://www.pcpro.co.uk/blogs/?p=7795#comment-122284 It’s nice to see my poster there in your picture (nearest on the right). On a more serious note, I would caution against describing CUDA as being only used by ivory tower academics doing high level physics simulations. A reasonable minority of us were actually thinking of real world applications (My poster described retinal prosthesis). The crazy concepts being pioneered in academia will become the products, and thus the revenue streams of tomorrow. It’s nice to see my poster there in your picture (nearest on the right). On a more serious note, I would caution against describing CUDA as being only used by ivory tower academics doing high level physics simulations. A reasonable minority of us were actually thinking of real world applications (My poster described retinal prosthesis). The crazy concepts being pioneered in academia will become the products, and thus the revenue streams of tomorrow.

]]> By: John Melonakos http://www.pcpro.co.uk/blogs/2009/09/30/reports-of-cuda%e2%80%99s-death-exaggerated/comment-page-1/#comment-110056 John Melonakos Sat, 10 Oct 2009 23:51:35 +0000 http://www.pcpro.co.uk/blogs/?p=7795#comment-110056 Disclaimer: I work for AccelerEyes. I just wanted to point people to the work we're doing to make it super easy for MATLAB code to run on the GPU using CUDA underneath the hood (see Jacket at http://www.accelereyes.com). As other GPU APIs mature, we'll support them, but our requirement is that they be useful to scientists (i.e. at least containing FFTs, some BLAS, and have a good plan for doing LAPACK-level functions as well). CUDA has all of that and is the only option today for any of the above. It'll be exciting to see AMD and Intel jump into the mix as their libraries mature, but they are several years behind in terms of the libraries that scientists, engineers, and financial analysts need. Disclaimer: I work for AccelerEyes.
I just wanted to point people to the work we’re doing to make it super easy for MATLAB code to run on the GPU using CUDA underneath the hood (see Jacket at http://www.accelereyes.com). As other GPU APIs mature, we’ll support them, but our requirement is that they be useful to scientists (i.e. at least containing FFTs, some BLAS, and have a good plan for doing LAPACK-level functions as well). CUDA has all of that and is the only option today for any of the above. It’ll be exciting to see AMD and Intel jump into the mix as their libraries mature, but they are several years behind in terms of the libraries that scientists, engineers, and financial analysts need.

]]> By: Tim Smalley http://www.pcpro.co.uk/blogs/2009/09/30/reports-of-cuda%e2%80%99s-death-exaggerated/comment-page-1/#comment-108646 Tim Smalley Thu, 01 Oct 2009 18:48:28 +0000 http://www.pcpro.co.uk/blogs/?p=7795#comment-108646 Gary: C for CUDA isn't 'industry standard C' - Nvidia has never said that to me, or any other member of the press as far as I'm aware - but my understanding is that it is very similar to writing conventional C code. I've spent quite some time speaking to a number of companies (one whose office used to be nextdoor to bit-tech's, before we moved into London) who have ported their apps to CUDA. Every one of them said that the porting process requires very little effort on their part. As for Nvidia's PR machine doing its job and paying developers to use CUDA (not your accusation, but an accusation that has come out of AMD on multiple occasions)... the company next door to bit-tech was asking me if I could help them to establish a contact window in order to help debug some of their code. They didn't have any contact with Nvidia, but they witnessed big speed ups in their applications with a very simple port from C to C for CUDA. I asked another British dev, who works for a finance company in Chicago, if he'd consider porting their applications from C to DirectCompute or OpenCL, since they're standards supported by both AMD and Nvidia, but the answer was a resounding no. He has moved the app from C# to C++ to C on the CPU and seen some decent performance increases CPU-only. Those performance increases were easy to justify, but he said the effort required to take a C program and port it to "a graphics API" (his words, not mine) was too great and difficult to justify to his bosses when the effort required to move from C to CUDA was minimal in comparison, but offered effectively the same speed ups. Gary:

C for CUDA isn’t ‘industry standard C’ – Nvidia has never said that to me, or any other member of the press as far as I’m aware – but my understanding is that it is very similar to writing conventional C code. I’ve spent quite some time speaking to a number of companies (one whose office used to be nextdoor to bit-tech’s, before we moved into London) who have ported their apps to CUDA. Every one of them said that the porting process requires very little effort on their part.

As for Nvidia’s PR machine doing its job and paying developers to use CUDA (not your accusation, but an accusation that has come out of AMD on multiple occasions)… the company next door to bit-tech was asking me if I could help them to establish a contact window in order to help debug some of their code. They didn’t have any contact with Nvidia, but they witnessed big speed ups in their applications with a very simple port from C to C for CUDA.

I asked another British dev, who works for a finance company in Chicago, if he’d consider porting their applications from C to DirectCompute or OpenCL, since they’re standards supported by both AMD and Nvidia, but the answer was a resounding no. He has moved the app from C# to C++ to C on the CPU and seen some decent performance increases CPU-only. Those performance increases were easy to justify, but he said the effort required to take a C program and port it to “a graphics API” (his words, not mine) was too great and difficult to justify to his bosses when the effort required to move from C to CUDA was minimal in comparison, but offered effectively the same speed ups.

]]> By: Gary Silcott http://www.pcpro.co.uk/blogs/2009/09/30/reports-of-cuda%e2%80%99s-death-exaggerated/comment-page-1/#comment-108529 Gary Silcott Wed, 30 Sep 2009 23:56:13 +0000 http://www.pcpro.co.uk/blogs/?p=7795#comment-108529 Disclaimer: I work for AMD. I feel obligated to comment since its obvious the Nvidia PR machine has done its jog - the idea that CUDA is somehow “industry standard C” is pure fallacy. If you speak directly to actual developers they know it's not accurate, but marketing has a different standard. If CUDA were “industry standard C” that would be 2 simple things: 1) you could take a CUDA program and compile it and run it with another standard C compiler; and 2) you could take a standard C program and compile and run it with the CUDA compiler: Actually you can’t do either. CUDA is not standard C. Disclaimer: I work for AMD.
I feel obligated to comment since its obvious the Nvidia PR machine has done its jog – the idea that CUDA is somehow “industry standard C” is pure fallacy. If you speak directly to actual developers they know it’s not accurate, but marketing has a different standard. If CUDA were “industry standard C” that would be 2 simple things: 1) you could take a CUDA program and compile it and run it with another standard C compiler; and 2) you could take a standard C program and compile and run it with the CUDA compiler: Actually you can’t do either. CUDA is not standard C.

]]>