None of the GeForce cards support Maximus at all. You must get a Quadro in order to use a Tesla card. And even then, for Premiere Pro, overall performance with even a Quadro 6000 and a Tesla C2075 is slower than a GeForce GTX 570, let alone a GTX 680.
Thanks Rj, how do you know that a Maximus config of, say, Q6K+Tesla, will be slower than a GTX-680, and if so - in what and by how much?
I need something tangible.
First off, the Quadro 6000 is actually a GTX 470 with a 384-bit (instead of 320-bit) memory bus. And I believe that I saw one result with that combo in the PPBM5 results list on the PPBM5 site. It proved no faster than a reference-clocked GTX 470 in those tests.
Are you saying that a Q6K+Tesla will not be any faster than a Q6K alone?
If so, is that a sure thing, or just guessing?
If anything, a Quadro 6000 by itself is actually a bit slower than a GTX 470 because the Quadro 6000 is clocked at lower speeds than the reference GTX 470. And adding a Tesla card does not add a sufficient number of cores to fully offset the deficiencies of the Quadros (which are essentially underclocked GeForces with a hige heatsink on them).
Harm, much appreciated - this answers the 1st, and partially, the 2nd question.
Let's assume the client wants a Quadro - 5K or 6K, and does not want a GeForce. So with a Quadro as a primaty card, does Maximus offer measurable benefits outside of smoother playback? If so, are there benchmarks to prove it?
P.S. Adobe and Nvidia tout 9x performance boost with Maximus - and looks like that only happens:
- with Q2K and to a lesser degree, with Q4K and up,
- on some quite strange timelines with 27 effects applied all at the same time. No appreciable benefit with "just" 15 effects.
- for FPS, i.e. smoother playback primarily (see Dennis' article) - does not affect rendering and encoding speeds much. In terms of realtime playback and scrubbing quality, Dennis does see significant benefits on four AVCHD clips.
So I just want to get to the bottom of that Maximus thing.
You have to keep in mind that with a single Q5/6K card, all MPE accelerated business is handled by the Q5/6K card, using all the available CUDA cores on those cards. When you move to a Maximus configuration, by adding a C2075 Tesla card, the CUDA cores on the Q5/6K cards are bypassed and only the CUDA cores on the Tesla card are used for MPE acceleration. It is as simple as that.
PR was never able to use multiple GPU's and effectively adding a Tesla card to create a Maximus environment only disables the Quadro card for hardware MPE and leaves it up to the Tesla card. The remaining function for the Quadro card is to steer the monitors. Hence my reservations about this Maximus solution, unless you need 10 bit output to your monitors.
Admitted, I do not know how this works with AE, since AE can use multiple GPU's, so there it may be advantageous.
The nine times performance improvement claimed in the article you linked to are about the same or less than we find in the Benchmark Results.
It is my conviction that memory bandwidth is the overriding factor that determines MPE performance and up to now our benchmark results support this.
If you compare the memory bandwidth of the various cards, you see this pattern emerge:
Video card Memory bandwidth Memory Bandwidth in Maximus configuration with a Tesla C2075 Quadro 4000 89.6 GB/s 144 GB/s Quadro 5000 120 GB/s 144 GB/s Quadro 6000 144 GB/s 144 GB/s Tesla C2075 144 GB/s 144 GB/s GTX 680 192.2 GB/s
Not possible, so it is limited to 192.2 GB/s
This makes it obvious that the fastest solution is the GTX 680. If opting for a Maximus solution, it makes no sense to choose a Quadro 6000 over a 4000, unless you need the memory size of the 6000.