Unified Memory in CUDA 6

[Michael H.W. Weber]

So far, RNA World does not profit from GPU computations, so at present, it makes no sense introducing these with the current applications.
However, future applications in the fields of molecular docking and molecular dynamics simulations would certainly profit from GPUs and then we should indeed employ them.

Michael.

[nico]

CUDA 6 cannot change the hardware. The problem is the slow PCIe bus between GPU and CPU. The current application needs to communicate a lot with the GPU which makes it inefficient to use the GPU at all.

[MReed]

As you're surely aware, the extreme power of a GPU comes by using parallel computing units, and quite massively so.
However, there are applications, which do NOT work, if parallelised and as far as I understand the current Infernal application, it is one of them. Therefore, CUDA will not help increasing performance in any significant way.

Either way, I am sure the development team of the RNAW-application will gladly employ the use of GPUs, if and when the developers of Infernal present a parallelisable version.
Until then, it's nice to see that there's a development in the field of GPU-processing, but even repeatedly posting the link to whatever paper there is, will not, in any way, speed up that process. In fact it might even hinder it, due to our development team having to give the same answer time and again.

[nico]

Zarck: Your diagram says it: "Theoretical GFLOP/s". The problem is always to get the GFLOPS with a real world application.
Don't get me wrong. CUDA6 is a huge step for developers and it will most certainly benefit the development of better applications. But as said above: It cannot change the speed of the PCI bus, which is the bottle neck for some applications like Infernal.

MReed: Impressive English! Where did you learn that?