Hi everyone,

I am going to by a "little" cuda capable card for my pc. But I dont want to spend much mony on the new card. So my prefered cards are the mentioned ones from above. My question is (and I already searched here in the forum) how much more performance the 9600 series really has. I talked to some guys in the stores and they say, that the 9500 and the 9600 have little performance differences in cuda applications. So I want you to ask how much more performance I would get if I buy a 9600gso instead of a 9500gt or a 9600gt instead of a 9600gso.

thanks for your answeres in anticipation :-)

p.s.: I hope my english was not too bad...

So I want you to ask how much more performance I would get if I buy a 9600gso instead of a 9500gt or a 9600gt instead of a 9600gso.

If your PS is more than 300 watts, do not even consider the 9500GT. 32 shaders will work okay, but some longer work will require more than 2 days by a few hours meaning that some "babysitting" will be occasionally required.

The 9600GT is a good card and will run, with stock clocks, around 24 hours per workunit with its 64 shaders (longer workunits a few hours more, shorter workunits a few hours less).

There are two versions of the 9600GSO. Those with 512mb or 1GB of memory are the newer version based on the G94 chip (same as the 9600GT) with only 48 shaders. It is slower than the 9600GT overall, including CUDA. The older version of the 9600GSO (with either 384mb or 768mb memory) with the G92 chip has 96 shaders and is the fastest of all the cards you mention for CUDA. Stock clocked cards (1375mhz shader) will run most workunits in less than a day (as little as 15-16 hours on the shortest workunits); my factory OC card (1700mhz shader) runs those short units in 13-14 hours. One thing to note, the 8800GS is also available new and is essentially the same card as the 96 shader version 9600GSO (NVIDIA renamed them as a marketing ploy to clear old stock).

One last thing to note...The older 9600GSO, while having more shaders and faster CUDA, is more or less equal to the 9600GT in many other applications and somewhat slower at times. This is largely due to the different bus widths (192-bit for the GSO vs. 256-bit for the GT). So, if you will be doing other GPU intensive activities, the 9600GT may be a better overall performance option.

First of all thanks a lot for the detailed answers :-)
I think my opinion will benefit from those.

But I´ve got a little fear from "babysitting", possibly I dont understand the word right. Does that mean, that WU´s cant be stopped while they are working? Usually I shut down my PC at night and the calculations continues the next day. Is this also possible with GPU-calculations?

If thats the case there would be no Problem having a slower card wouldnt it?

By "babysitting" I mean you cannot simply install the card and let GPUGRID run. For example, let's assume you have a dual core PC. GPUGRID will download 2 workunits because allocation of work is by number of cores rather than number of GPU's (...this might change in the future given the new BOINC client changes in work-fetch policy). Thus, there will be times when you will download 2 of the longer workunits (or even worse, 3 or more in a row). Since these longer workunits will exceed two days each in runtime on a 32 shader card, you will exceed the 4-day completion deadline set by the project. If you are over by only a little (basically less than a few hours--the processing speed of the fastest NVIDIA card at that time) you will still get credit for the late work. However, if the workunit is reassigned and completed before you return it, then you get nothing. Thus, you will need to "babysit" (keep an eye on the card regularly) to make sure you are not stuck in a loop of always being behind and getting no credit (i.e. - you will occasionally need to abort a workunit or two). Cards with 50+ shaders or so do not run into this problem.

ah ok - I see what you mean.
But the WU´s are stoppable are´nt they?

Yes. You can certainly pause the WU's. The issue you encounter is having WU's pending / running that have no opportunity to finish on time. These will be resent out after the 4-day timeframe to another user, who will then work on it. The first person to respond gets the credit. The babysitting is that you might need to abort some WU's if they have no opportunity of finishing.

I have 2 8800GT's and one 9500GT. I have no issues with this even with my Quad-core because the 8800GT's take about 12 hours per WU (compared to about 40 for the 9500GT), so they will process the extra WU's within the timeframe (4 WU's at once, 3 running and 1 pending).

You can probably have a 9500GT with a dual-core and not generally run into issues unless you game a lot or otherwise use the gpu for something else. This is due to the fact that the # of WU's sent is tied (currently) to the # of cpu's you have in your system (I know, a little backwords, but gpu-computing is in its infancy). If you have a Quad, then you definitely will run into issues -- you'll have 1 WU running and 3 pending, two of which will definitely have to be aborted to make the 4-day timeframe.

Hopefully this helps.

I understand now the problem.
Thanks for your help :-)

Due to the fact, that I usually "babysit" the WU´s of other project (simply because I often sit in front of my pc ^^) there should be no problem with babysitting for me.

But for now I think a 9500 is the better choise for me because of financial issues. And in a half year or so I will buy a completely new PC with a new better card... and so i wont waste money into hardware which will be replaced ...

So. Thanks to all of you ;-)

But for now I think a 9500 is the better choise for me because of financial issues. And in a half year or so I will buy a completely new PC with a new better card... and so i wont waste money into hardware which will be replaced ...