OctaneRender® pre-Beta 2.41 (win x64) [OBSOLETE]

[abstrax]

to samCameron

Further to my post above.

This is a quote from the Annandtech site i linked to regarding testing of the GF104 - Fermi chip as used in the GTX460

"Due to the fact that NVIDIA added an additional block of CUDA cores to an SM without adding another warp scheduler, the resulting superscalar design requires that the card extract ILP from the warps in order to simultaneously utilize all 3 blocks of CUDA cores.

As a result the range of best case to worst case scenarios is wider on GF104 than it is GF100: while GF100 could virtually always keep 2 warps going and reach peak utilization, GF104 can only reach peak utilization when at least 1 of the warps has an ILP-safe instruction waiting to go, otherwise the 3rd block of CUDA cores is effectively stalled and a GTX 460 performs more like a 224 CUDA core part. Conversely with a total of 4 dispatch units GF104 is capable of exceeding GF100’s efficiency by utilizing 4 of 7 execution blocks in an SM instead of 2 of 6.

Or in other words, GF104 has the possibility of being more or less efficient than GF100."


This then may be a possible explination on why the GTX460 - using the GF104 GPU design ,although appearing to use all the cores is still no faster using Cuda 3.2 and is in fact only as fast an old GTX260/280 card and a lot slower than the GTX 465/470 cards which used the older GF100 GPU.

Perhaps someone at Refractive could comment.

Yes that all sounds correct to me, except: A GTX 460 is still heaps (more than 2x) faster than a GTX 260. I have got one GTX 260 at home and on the chess scene I get ~4.15 Msamples/s while on the GTX 460 I get ~8.75 Msamples/s

In general there seems to be a misconception regarding cores and stuff: Usually it's not possible to keep all cores busy all the time. Everytime there is a branch in the code, it is possible, that part of the shaders processors (/cores) used for one block have to wait. The problem with the GTX 460 is, that the block size is larger and therefor a larger number of cores potentially gets stalled. I guess, this what we see here. BUT as I said before: GTX 460 is still a lot better than a GTX 260.

Cheers,
Marcus

[enricocerica]

Octane with 3.2 Cuda driver ignores the external GPUs ! So, external boses like Cubix seems to be not usable with Cuda 3.2 !

Please, Cubix users test and confirm !

My cards located in the Cubix box are well recognized and I have no problem with them. I use Winxp64 with the recommended drivers.

[ROUBAL]

@Enricocerica : The cards in the Cubix are seen in the Octane manager, but if I select a card other than the one in the computer (used for display), Octane doesn't load the ocs or obj in its memory and crashes when I try to close it.

My Internal GTX260 is sufficient to launch Octane and render slowly a small scene for testing purpose, but the UI is very few responsive.

these two days, I was in panic mode and I tested night and day, doing dozens of hard reboot and drivers install-uninstall. This evening, I had to go out for pruchasing food for the week, and I thought about the problem.

Maybe, it is a matter of misunderstanding : When one says non multi GPU support, does it mean that you have to manage the GUI and render with THE SAME GPU, or does it mean that only one GPU must be selected in the Octane Cuda manager ?

These two possibilities are very different, and maybe I'll just have to wait for the multi GPU release ? I would really know if I just misunderstood the way Octane works with GPUs, or if there is really a problem with Cuda 3.2 for external graphic cards.

I would really like an explanation from Refractive, because I'm really worried : if Cuda 3.2 wouldn't support External GPUs, I couldn't use my Cubix box and Octane releases from now ! "Panic mode ON"

[abstrax]

I reinstalled 260.99 and didn't switch on my Cubix. I launched Octane 2.41 and loaded a very small ocs with only the head of a character I'm working on : It worked, but with only my GTX260, it is not usable with two big screens !

I just saw that my reported bug about the horizontal bar when subsampling is enabled has not been fixed.

Sorry, I forgot it. Will do it today.

Then, I rebooted with the Cubix. I had to reinstall the 260.99 for each of my two GTX480, and reboot each time.

When done, I launched again Octane 2.41 : First it opened but refused to load an ocs, and frozen the computer requiring a hardware reboot.

Sometimes Octane doesn't even open, and freezes the computer when launching, and sometimes opens but doesn't load ocs or obj ! The mesh node is created, but clicking onit has no effect : the memory is not filled.

All other softwares work fine with 260.99. Finally, I think that I found out what is really going wron :

I tried again with the Cubix Xpander OFF, and it works with only my GTX260. Initially, I thought that it was because of not supporting multi GPU, so I have changed the Cuda preferences in Octane to keep only one GTX480 for computing : same behavior with only one GPU : Doesn't load ocs or obj files.

It is worse than expected :

Octane behave like ignoring the external GPUs ! My Cubix with its two GTX480 seems to be not usable with Cuda 3.2 ! It is a nightmare !

I have tried to switch back to Octane 2.3v5 : Same trouble ! So it is really due to the driver with Cuda 3.2 not supporting external GPUs !
2.3v5 with only the weak internal card, works (very slow and unresponsive, obvioulsly).

I bought the Cubix because my workstation can't host these cards. I'll hope that this problem will be solved soon ! Non multi GPU support is one thing, but if you can't access to your external GPU even if selecting only one, it is not the same !

Please, I would like to get some reports from Cubix users.

Sorry, Guys... I still can't use this release.

Ok. Could you please take one card out of the expander and try again?

Regarding multi-GPU: Octane indeed runs on several GPUs, but as you have already seen: The actual GPU usage is somehow throttled, the more cards you use and in the end it even gets slower than with one GPU only. That's why we will release a CUDA 3.0 build today. Please try with that one again and let us know, if it works better.

Cheers,
Marcus

[abstrax]

@Enricocerica : The cards in the Cubix are seen in the Octane manager, but if I select a card other than the one in the computer (used for display), Octane doesn't load the ocs or obj in its memory and crashes when I try to close it.

My Internal GTX260 is sufficient to launch Octane and render slowly a small scene for testing purpose, but the UI is very few responsive.

these two days, I was in panic mode and I tested night and day, doing dozens of hard reboot and drivers install-uninstall. This evening, I had to go out for pruchasing food for the week, and I thought about the problem.

Maybe, it is a matter of misunderstanding : When one says non multi GPU support, does it mean that you have to manage the GUI and render with THE SAME GPU, or does it mean that only one GPU must be selected in the Octane Cuda manager ?

These two possibilities are very different, and maybe I'll just have to wait for the multi GPU release ? I would really know if I just misunderstood the way Octane works with GPUs, or if there is really a problem with Cuda 3.2 for external graphic cards.

I would really like an explanation from Refractive, because I'm really worried : if Cuda 3.2 wouldn't support External GPUs, I couldn't use my Cubix box and Octane releases from now ! "Panic mode ON"

Don't panic yet. When we say no multi-GPU support, we mean no good performance on a multi-GPU setup. At work for example, I have a GTX 470 and GTX 460 and I can use both cards for rendering or only one. It always works, but when I use the GTX 470 + GTX 460 I get a worse performance with CUDA 3.2 than with a GTX 470 alone. It behaves as if the second card can start its work only, when the first one has finished, i.e. they are not running in parallel. We haven't found the reason for that yet, hence the CUDA 3.0 and CUAD 3.2 builds we will do.

Unfortunately I don't have a Cubix Xpander available at the office and thus can't do any tests. -> If someone else with a Cubix box can share some experience, it would be great.

Cheers,
Marcus

[ROUBAL]

Hi Marcus, I 'm happy to get a response. Forget my PM, I asked the same things.

Well, you give me some hope. As I have almost not had time to sleep or eat for two days, I'm not sure that I will open the Cubix to remove a card this evening. I'll do as soon as possible.

Reports from Cubix user would be welcome.

[enricocerica]

@Enricocerica : The cards in the Cubix are seen in the Octane manager, but if I select a card other than the one in the computer (used for display), Octane doesn't load the ocs or obj in its memory and crashes when I try to close it.

My Internal GTX260 is sufficient to launch Octane and render slowly a small scene for testing purpose, but the UI is very few responsive.

these two days, I was in panic mode and I tested night and day, doing dozens of hard reboot and drivers install-uninstall. This evening, I had to go out for pruchasing food for the week, and I thought about the problem.

Maybe, it is a matter of misunderstanding : When one says non multi GPU support, does it mean that you have to manage the GUI and render with THE SAME GPU, or does it mean that only one GPU must be selected in the Octane Cuda manager ?

These two possibilities are very different, and maybe I'll just have to wait for the multi GPU release ? I would really know if I just misunderstood the way Octane works with GPUs, or if there is really a problem with Cuda 3.2 for external graphic cards.

I would really like an explanation from Refractive, because I'm really worried : if Cuda 3.2 wouldn't support External GPUs, I couldn't use my Cubix box and Octane releases from now ! "Panic mode ON"

Ok I just made some more tests, I have a GTX275 and a GTX480 in the Cubix box. I can select one or the other card and it works fine, I can select both cards and it works fine too, abviously without performance increase as announced, so I see no problem with multi card in a Cubix box.

[abstrax]

Ok I just made some more tests, I have a GTX275 and a GTX480 in the Cubix box. I can select one or the other card and it works fine, I can select both cards and it works fine too, abviously without performance increase as announced, so I see no problem with multi card in a Cubix box.

Thanks for the feedback. Could you do one more test: Try to run both cards in your box?

Thanks,
Marcus

[enricocerica]

Ok I just made some more tests, I have a GTX275 and a GTX480 in the Cubix box. I can select one or the other card and it works fine, I can select both cards and it works fine too, abviously without performance increase as announced, so I see no problem with multi card in a Cubix box.

Thanks for the feedback. Could you do one more test: Try to run both cards in your box?

Thanks,
Marcus

I've not been clear enough , I also tested to run both cards and it works.

[abstrax]

I've not been clear enough , I also tested to run both cards and it works.

Sorry, obviously I'm not awake yet... Thanks for the info anyway.

Cheers,
Marcus