Just when we thought we might get around to moving to 3.2 the green team announces 4.0! Well.. actually a pre-release version will come soon (4th Mar) and there's been substantial progress. This should definitely make developing easier. Not so sure about backwards compatibility, though.

MrS
____________
Scanning for our furry friends since Jan 2002

Not so sure about backwards compatibility, though.

MrS

400 series should be fine. There is no chip change as yet.

With the advent of CUDA4, we will drop cuda2.2 and support only cuda3.1 and cuda4.0.
This is when we will come out with cuda4 which is not urgent.

gdf

Not good news for CC1.2 cards; they dont fair well with 3.1.
Why not just stick with 2.2 (up to cuda3.0 drivers) and move from 3.1 to 4.0?

Because fermi cards without the latest driver would not work.

gdf

Fair enough, there are lots of people that don't have the latest driver - perhaps around half would get caught out, and that's a lot of lost potential.

I have expected this for some time, so thanks for the early warning. I needed a shove to move on some of my GT240's anyway; one GTX470 could now get more credit/do more work than all 6 of my CC1.2 cards, and for about half the Wattage.

At present I am seeing around 70% GPU utilization for the long tasks, but perhaps this will change with tuning.
Most normal length tasks are now using the improved routines so it's common to see over 90% utilization. Obviously your exact GPU utilization depends on the specific task, the app and other parameters such as your CPU, it's usage, GPU usage and your choice of optimization settings.

Given that the 6.12app is to be pulled and the poor performance of the GT240 running the 6.13app, at this stage I would say there would be little point getting a GT240. For those that have one, they will be good right up until we have to use the 6.13app. Cuda 4.0 is still in Beta so it might be a while yet.
I wouldn’t anticipate any future performance improvements on the 6.13app for CC1.2 cards, or following the use of cuda 4.0.
While the GT440 and GT440-OEM should work, they will be slow and would not feature too highly on the recommended list. If someone posted actual results then we would know the performance to expect, but we are certainly talking short tasks only.

I agree. It’s unfortunate that no-one has posted the results of a GT440. As is I would not recommend getting one. So unless your systems could take GTS450 or GTX550 Ti cards (106 and 110 TDP) then I guess you should play the waiting game; these planned changes along with the CC2.1 design issues mean it’s not a good time to buy an entry level crunching GPU for GPUGrid.
This month AMD is expected to launch its HD 6990 Antilles and NVidia is expected to launch their dual Fermi rival, so NVidia’s attention presently lies elsewhere.

While GPUGrid could really do with an entry level 128 cuda core Fermi and a mid range 256 cuda core CC2.0 Fermi, I can’t see that happening before the next generation of Fermi’s.
It is still possible that a GT540 (or similar) could turn up with 3 SP’s and good clocks; the 500 series is lacking variety. Cuda 4000 could also bring some performance improvements that would make the GT440 an obvious choice, and sooner or later manufactures will start messing with the 500 series designs, so you might see a green version of a GTX550 Ti with a TDP of about 80W, or a revved up/3SP version of a GT440/GT540. The 41W difference between the 65W GT440 and the 106W GTS450 is quite large, so I would not be surprised if something arrived using up to 85W, though this would require using a power connector.