Using simultaneously Fermi and non Fermi cards ?

[ROUBAL]

Hi, I'm still trying to save money in the goal of purchasing better graphic cards at the end of the year.

As it is really difficult at an affordable cost (I mean without purchasing a tesla C2050 or C2070)
to get both high rendering speed (for animations for example) and a large amount of memory (for high resolution billboards and very detailed scenes), I was thinking of purchasing a GTX 480 (Fermi) and a Tesla C1060 (Not Fermi, but the only card I could afford with so much VRAM).

The GTX 480 has 480 cores and 1.5 GB of VRAM.
The Tesla C1060 has 240 cores and 4 GB of VRAM.

In theory, using simultaneously the GTX 480 and C1060 in my Cubix box, I would get 720 cores and 1.5GB of VRAM when I need speed, and 240 cores and 4 GB of VRAM (Tesla C1060 only) when I need high memory space.

This is the theory, but in fact I don't know if Fermi and non Fermi cards can operate together, and if yes, it there is not a slow down due to the slowest card.

It is important, because not using the power of the 240 cores of the Tesla card when working on a simple scene, would be a waste of ressources.

If someone has some experience about mixing the two technologies, it would be helpful !

Thanks in advance.

Philippe.

[timbarnes]

I have tried to use my GTX 470 at the same time as the Quadro FX 580. It was not a success, because the 580 is too slow, and holds the 470 back. The 470 alone is faster (and I get a usable display running on the 580).

I don't know the relative performance of the two cards you're looking at, however. In general, I gather it is best practice to use similar cards if using two or more for Octane simultaneously.

[Nostromo]

Yeah. Also, when using two card, there's other factors that play too like, for example, switching openGL context from one card to the other (to compose the full image) and there's no number for that unfortunately. Fermis seems to be very good at switching contexts (that's why they provide good scalabilty) while Quadro's seem to be a little slower to do this.

I think indeed using 2 equivalent architecture is indeed the best to avoid deception.

/M

[ROUBAL]

Thanks for the answer.

Well, as I can't afford two Tesla, even old C1060, I think that the most reasonable solution remains two GTX 480. Currently, I reach the limit of my GTX 260 (877MB).

One Tesla C1060 only would be convenient in term of memory amount and not much more expensive than two GTX 480, but almost as slow as my GTX 260.

1.5 GB is not much, but for still pictures I can remove some hidden parts... and I will get much speed for animation... (when we'll have a faster obj exporter !)

[n1k]

For animation it shouldn't be much of a problem having fermi and non fermi cards rendering same animation. Lets say you have 600 frames to render. You can run two instances of octane and choose to render, for example, 400 frames on faster and 200 frames on slower card. That way you can have perfect scalability:)

Cheers,
n1k

[ROUBAL]

That's a weird idea ! But if it works, it is very interesting !

I wouldn't have thought that it was possible to open two instances of Octane...
mainly because of license managing with Octane live.

Have you tried ? As I have currently only one card in my Cubix, I can't make switching cards tests myself, except just opening an instance.

[Nostromo]

In theory it should work since the host is going to be the same.

/M

[havensole]

I haven't tried it on the new pre-2.3 version, but I've had it work with older versions of Octane. When I had 3 gpu's in my system at one point (2x gtx470's and a gts250) I would run 3 instances of Octane all rendering different things. I did this just to test how much I could abuse my system with Octane. never had an issue with it. I'll probably try it here with the pre-2.3 later today.

[kubo]

I didn't know that, hey that's a great tip, I thought my 260 was getting lazy, lol. Time to burn that sucker!

[radiance]

Hi guys,

an important fact, you cannot combine GTX200 and GTX400 cards during a render.
they are 2 different architectures, much like the difference between an intel CPU and a PowerPC RISC CPU.

cuda decides what architecture to use based on the first chosen GPU in your active gpu list, so if you have a GTX200 and a GTX400 type GPU in there, the GTX200 kernel will be loaded into the GTX400 and the application will crash.

there is'nt much we can do about this, it's a cuda limitation.

if it does, please report, but it should'nt.

Radiance