Hi, questions regarding benchmarks:
In the benchmarks article, Server 1 does not show a CUDA GPU, and yet there is a column for "Server 1: CUDA" results. What does this represent? Was this using the same Tesla C2070 "rocket" as Server 2? Or, a more modest GeForce video card? If different, then which?
Is there really a need for the Tesla C2070? What level of floating-point precision arithmetic does the transcoding/transrating algorithm use? If it only uses single-precision then as I understand it a cheaper GeForce GTX 580 actually outperforms the Tesla card. I'm only interested in transrating, if that makes a difference.
As I said, I want to do transrating of multiple streams. Whatever it is that Server 1 is using for its benchmark appears to be similarly powerful to Quicksync. If I get a Z68 motherboard with whatever that CUDA card is, and then run half my jobs on CUDA and half on QuickSync, do I get better CPU utilization? Or, does the CPU spend about as much time servicing each job regardless of the GPU so it doesn't help?
How do you tell when you are reaching saturation on the capabilities of the GPU? CPU utilization may still be relatively low, but if you've exceeded the capacity of the GPU then presumably something starts to break. Is there a way to measure GPU utilization? If not, what happens if you push too hard? Do you get error messages? Or, do the streams just start to look bad?
Is there any perceived difference in quality between CPU vs. Quick Sync vs. CUDA?