now I am getting ready to implement CUDA into my physics library.
I am reading the programmers guide for the n time to make sure I understand well the architecture.
I must said the more I read the more I like it in many areas, and not so much on others.
I have two questions.
I read somewhere that it is possible to plug more than one card on the same machine, but the manual say that the cards must be identical.
I am confused, my motherboard have three PCI express slots, and I am wandering what would happen if a plug a g8600, or g8500 in the primary PCI express and a g8800 to another slot.
The reason I ask is that I am interested in using the card as a math co-processor only, for acceleration of all kind of physics simulation: rigid bodies, animation, rigid particles, sph fluids, cloth, etc
I read somewhere here in the forum that when the final CUDA SDK is released it will work on all g80 series of hardware but that for now only gtx8800 and gts8800 are supported. Is this correct?
Ideally what I want to do is to used one g8500 as a standalone co-processor, but I do not want to jeopardize my machine plugging two different cards and having catastrophic results.
For what I can see in the API, there is a function to queries the hardware, I do not see why they have to be identical.
My second question is about the design and marketing of the g8500 and g8600, for what I can see they are targeted to the mass consumer and entry level user.
I went to a store on the weekend and saw the g85 for $149.99 with 256 meg of ram.
I assume they do not have the bandwidth and the floating point throughput of the g88s
so I do not understand why it is that they are PCI express only, or at least I did not see any AGP versions.
The following is a speculation and a wish.
For what I understand there are more than 20 million PC in the US but I am willing to bet that more than half of those PCs are older generation from the PCI an AGP time.
I understand that the top of the line cards needs the high bandwidth, but I do not see how the PCI express bus will help on cards that are targeted to older PC.
On the other hand an AGP version of the same card will allow for so may users to update to better a graphics solution with the option of better overall performance on application that take advantage of the hardware for things other than graphics, and all for a very reasonable price. After all isnt that the reason CUDA was created?
Finally and this is a wish, why no make PCI version of the g8500, there are lots game companies and middleware developers that could get good used of hardware acceleration, yet not all these companies can afford expensive commercial licenses or do not use exclusive hardware solution, because they target their game to mass market. But if there was the possibility to use acceleration on low end hardware…, I do not know but I got the hunch that the sales will larger than anybody can possible imaging.
My impression is that CUDA was made to break the saturation of graphics hardware, but it is crippled by the bus selection for all versions, the way it is now is only usefully for high end users that will get then anyway.
I hope I am not out of line here speaking nonsence, as I usually are.