Official Statement from Otoy towards GTX 980/970

[smicha]

Tutor,

I wish I had opportunity to help you with testing the 980 - I just simply don't have access to them.

What might be interesting is that the more vram is used (for large scenes) the more likely throttling will occur. I remember I was rendering 4.5GB scene and although my titan is super stable at 1200 MHz I had to set it to 1000 MHz due to throttling - power usage was jumping up and down from 80 to 90%.


Glimpse,

I just read your article about noise levels. Great! What is important for me (I did some tests) is that 'path term power' set to 0 slows down PT mode but gives significantly better, smoother shadows.

[smicha]

Ok, so here are the results with Alpha Shadows Off:
Scatter.ocs - Octane 2.12.1
DL - 8.11 Ms/sec, 1000 samp/px - 00:01:05

Something is wrong...this would make my single 780 6x faster.

[pBarrelas]

Ok, so here are the results with Alpha Shadows Off:
Scatter.ocs - Octane 2.12.1
DL - 8.11 Ms/sec, 1000 samp/px - 00:01:05

Something is wrong...this would make my single 780 6x faster.

Yep, I repeated the test and it's giving me a slower result than with alpha shadows on...weird!

[Tutor]

Tutor,

I wish I had opportunity to help you with testing the 980 - I just simply don't have access to them.

What might be interesting is that the more vram is used (for large scenes) the more likely throttling will occur. I remember I was rendering 4.5GB scene and although my titan is super stable at 1200 MHz I had to set it to 1000 MHz due to throttling - power usage was jumping up and down from 80 to 90%.


Glimpse,

I just read your article about noise levels. Great! What is important for me (I did some tests) is that 'path term power' set to 0 slows down PT mode but gives significantly better, smoother shadows.

Like you, I don't have any GTX 980s to test myself and it does seem completely reasonable that rendering larger scenes are likely to induce throttling and that down clocking a GPU can reduce throttling. However, there's art and science at play here because throttling, itself, causes core down clocking - kind of like me, as a child, giving myself a spanking when I had done wrong to avoid my mother/father doing it.

[Tutor]

Ok, so here are the results with Alpha Shadows Off:
... .

Thanks for the information/update.

[Tutor]

Hi Tutor,

I have similar questions as pBarrelas. The tests I have been running have been for both Octane and Blender Cycles.

I tried running Octane 1.2 but it doesn't recognize the 980 so I used 1.55 instead. This is using the benchmark trench scene completed to 1000 s/px.

DL - 19.63 Ms/sec, 1000 samp/px - 31 sec.
PT - 5.96 Ms/sec, 1000 samp/px - 1 min. 41 sec.
PMC - 4.4 Ms/sec, 1000 samp/px - 2 min. 18 sec. (this has the alpha shadows on, turning them off errors out Octane and I had to restart the program)

These are the times for version 2.12.1:

DL - 15.07 Ms/sec, 1000 samp/px - 40 sec.
PT - 4.16 Ms/sec, 1000 samp/px - 2 min. 28 sec.
PMC - 3.62 Ms/sec, 1000 samp/px - 2 min. 46 sec.

These numbers should change once we get to use the next version of Octane. Marcus has already been able to improved the speed of the PT kernel to within 10 to 20 percent of the speed of a Titan Black.

Thanks for the information/update.

[Tutor]

... .My own theory is that it's about memory usage - once the scene is loaded Octane seems to do relatively little with it (don't know about animation in this regard). ... .

Thanks greatly for pointing this out - when the scene is being computed CUDA, not Octane, is doing the heavy lifting. Does anyone know whether CUDA been tweaked for Maxwell 980s?


Thank you pBarrelas and Grimm,
It's good to get a closer comparison of those two GTX 980s.

Ok, so here are the results with Alpha Shadows Off:
Scatter.ocs - Octane 2.12.1 (dual 6-pin power GTX 980)
DL - 8.11 Ms/sec, 1000 samp/px - 00:01:05
PT - 4.11 Ms/sec, 1000 samp/px - 00:02:08
PMC - 3.06 Ms/sec, 1000 samp/px - 00:02:52

Hi Tutor, ... .
These are the times for version 2.12.1 (dual 8-pin power GTX 980):
DL - 15.07 Ms/sec, 1000 samp/px - 40 sec.
PT - 4.16 Ms/sec, 1000 samp/px - 2 min. 28 sec.
PMC - 3.62 Ms/sec, 1000 samp/px - 2 min. 46 sec.
... .

,

Assuming that Grimm had turned Alpha Shadows Off also, and neither Grimm nor pBarrelas had changed their GPUs clock settings, then the Classified-like (c) GTX 980 is about 62% faster than the original (o) GTX 980 when it comes to rendering the DL scene [akin to cartoon type animations (rasterizations)] and for the PMC scene about 4% faster. When it comes to rendering the PT scene, the oGTX 980 renders about 16% faster than the cGTX 980. So I believe that for heavy rendering chores requiring raytracing in Octane, a significantly underclocked cGTX 980 may improve its rendering performance because it appears to be throttling under CUDA supported ray tracing operations in Octane. Also, it's having a higher TDP is not a cure to this problem in Octane. So, in the end, it just might be that, for whatever reason (such as driving content creators to higher priced cards or simply concentrating on improvements in OpenCL computing and neglecting CUDA this go around), Nvidia may not have sufficiently tuned these cards (or CUDA for these cards) for heavy duty CUDA ray tracing computing. However, note that in Furry Ball GPU rendering, a single GTX 980 excels above all other single GTX cards at cartoon type animations (rasterizations) and achieves high scores not far behind a single GTX 780 Ti or Titan card in raytracing [ see, e.g., http://www.aaa-studio.cz/furrybench/benchResults4.php and http://furryball.aaa-studio.eu/aboutFur ... marks.html ]. Thea's GPU renderer doesn't yet appear to be experiencing problems with the GTX 980 [ http://www.thearender.com/forum/viewtop ... it=GTX+980 ]. So even tho' the problem might appear to be more likely Octane related, the problem appears to exist with ReShift3d users also { https://www.redshift3d.com/forums/viewthread/2768/ ]. So now I'm completely stumped and will not purchase a Maxwell card until the problem is fixed, unless, of course, I ran Furry Ball or Thea Render. However, the biggest disadvantage of Furry Ball and Thea is that neither of them fully support C4d as well as Octane currently does.

[abstrax]

All the discussion about the 980s is mute, because the version you currently have available (2.12.1) is running significantly slower on a 980 than what I could achieve so far after tweaking Octane.

As mentioned here before, it's still slower than a Titan Black (and with that a 780 Ti), but quite a bit better than what you can play with right now. In other words: Please don't waste your time benchmarking the GTX 980 with an obsolete version and just wait. It's still work in progress.

In case you might wondering: GPGPU can be extremely sensitive to how code is written, memory accessed etc.. Octane is mostly memory bound and because of that, memory throughput, cache architecture etc. have a large influence on the speed you get out of it in the end. Unfortunately, this area changed a lot from Kepler to Maxwell, which is why some changes in Octane are required. And that's what I'm working on right now.

[pBarrelas]

All the discussion about the 980s is mute, because the version you currently have available (2.12.1) is running significantly slower on a 980 than what I could achieve so far after tweaking Octane.

As mentioned here before, it's still slower than a Titan Black (and with that a 780 Ti), but quite a bit better than what you can play with right now. In other words: Please don't waste your time benchmarking the GTX 980 with an obsolete version and just wait. It's still work in progress.

In case you might wondering: GPGPU can be extremely sensitive to how code is written, memory accessed etc.. Octane is mostly memory bound and because of that, memory throughput, cache architecture etc. have a large influence on the speed you get out of it in the end. Unfortunately, this area changed a lot from Kepler to Maxwell, which is why some changes in Octane are required. And that's what I'm working on right now.

Those are good news! We'll benchmark when the new version comes out.
It terms of hardware, accordingly to Nvidia, a GTX 980 should be faster than a Titan Black or GTX780Ti:
http://www.geforce.co.uk/hardware/deskt ... erformance
Too bad that isn't happening also here.

[Rikk The Gaijin]

Good news indeed! Thanks Abstrax!
Now, if you could also give us a version that supports the Oculus DK2, that would be great...