FAH-Addict

nVidia's latest GPUs have now arrived, along with information on their clock frequencies and final prices. Here is a quick summary before we move on to the technical stuff:


GTX 470



GTX 480



After a series of rather premature tests published on different parts on the Internet, one hesitates between uncontrollable laughter, and a psychotic depressive breakdown. There has been a great deal of nonsense written, which could consume many pages of A4.

That's hot

We shall begin on the hardware side. There are two noteworthy things here: it runs very hot, and it's a hungry beast. Many testers have found that the average operating temperature stands between 90 and 95°C, in both their standard tests and in their fully 3D ones, i.e. with GPU folding. Many folders have reported card lifetimes of a mere 15 days with certain cards when folding 24/7. Obviously, it would be advisable not to just throw the cards in the trash, but rather to wait for some official word before deciding what to do next.

Now for the Folding@Home tests. We shall forego the otherwise obligatory criticisms that are levelled by some reviewers who repeat word for word the press release information. Instead, we will focus a little more on practical results.


PPD? Where?

Most testers have given results in ns/day format. We would remind everyone that the ns/day is not a reliable way of accurately measuring GPU folding performance, for three reasons: firstly, it varies greatly with the unit being processed at the time; secondly, it is not comparable from one client to another; thirdly and most importantly, it does not take into account the points credited for processed work (hence the use of the PPD measure).


Questionable test figures

Moreover, it remains unknown how the figures provided by nVidia were measured, and we do not know what setup was used to obtain the ns/day values. A similar approach was used for the GTX 280, and we can only advise taking these unreliable figures with a pinch of salt.


GPU3 clarification

Finally, in folding-related reviews of the cards, some testers have written that the GPU3 client is alpha, not yet distributed, and only works on nVidia cards. Fair enough. But they probably do not know all the details of the future GPU client. As previously announced, GPU3 is going to involve a dual migration: the replacement of Gromacs with OpenMM, and migration of proprietary technologies (CUDA and Stream) to a new unified technology, OpenCL.

This migration will happen in two stages. We will refer to the first stage as GPU2.5 for ease of explanation. This stage will involve the client utilising OpenMM, but at the same time retaining the use of CUDA. That is why the new OpenMM core (FahCore_15) will only work on nVidia cards at first. The second stage will lead to the more casual user friendly GPU3 which will migrate completely to OpenMM once enough ATI support has been achieved.

Moreover, contrary to what some have written, the GPU2.5 core utilising CUDA is compatible with all CUDA cards from the 8xxx series through to the 4xx GTX series. This is also true of current cores (FahCore_11 and FahCore_14) which will also work without problems on the GTX 4xx.


Hope remains

To conclude the actual card review, we can estimate the practical gains in PPD relative to those reported by the other testers. If the gains from 50 to 70% observed on a GTX 480 over a GTX 285 are confirmed in practice, we can expect to see this card produce somewhere between 13,000 to 16,000 PPD. Be careful though with the power consumption, which has jumped about 40% between the GTX 285 and GTX 480.