Just to summarise, I'm in the process of doing something crazy, as I'm wanting to pair a modern GPU with a dual NetBurst Xeon from 2005.
The purpose of this exercise is to achieve the absolute maximum performance that I can from this system.
This is the Motherboard that I'll be using for this exercise: (It's a Tyan Thunder i7525 (S2676) fitted with two 604-pin Xeons @ 3.8GHz)
After discussions with "The Serpent Rider", I've reached a point where I'm considering these two GPU's: (Right click and select to open the image in a new tab if you can't see it clearly)
The ZOTAC 1080 Ti clearly has better theoretical performance, but the TITAN XP has significantly better memory throughput and more CUDA Cores.
So my question is the following: In the case where these GPU's are CPU bottlenecked to the extreme, will the difference in memory throughput (as well as the higher CUDA Core count on the TITAN) be an overriding factor when it comes to real world performance?
Remember that the goal here is to get maximum performance out of this dual NetBurst system, so please don't think as to how these cards would compare in the more modern type of system (i7) that was intended to run them.
Seeing that GPU's are SIMD-type systems, I'm personally of the opinion that the higher CUDA Core count would be a factor in this case. (Rather than the speed at which each individual core runs.) As you would accomplish more parallel calculations per instruction issued to the GPU by the CPU.
Seeing that the rate at which these CPU's can issue instructions the card are much lower than the desired rate, would greater parallelism not be a more desirable feature in this abnormal case?
Any inputs would be much appreciated!