Catmull Clarked Polylimit on 2 X Titan X 12GB?

[Gineus]

I have read several threads on polylimits. The info in them did not really help, it seems. One guy seems to have hit the roof at 40 million.

This is not critical, I am just interested what I can achieve with Octane 3.

I generally tried the polylimit by upping up the catmull clark in the object setting. I have used a card alone, priority on both or no priority and not used on one card.

The result is the same. Although I supposedly still have memory left, the render fails. It also does not recover after I lower the subdivision.

I am on sub 2 on a ZBrush model that is quite complex already.

I attached 3 files, I hope this is enough info about the situation and my system.

So the question is, is there an internal difference between polygons loaded vanilla and created through internal subdivision in Octane?

If there is no difference, what could be the reason for the polylimit not going to the end of VRAM memory?

I am using 3 alpha 6 (standalone) on 2x 12 Nvidia Titan X (Driver 364.72) in SLI Configuration
Intel Core i7 Extreme 5960X 3Ghz 20MB Cache (Vanilla not overclocked)
Gigabyte X99 Gaming 5P
64 GB Ram HyperX Fury DDR4 2133 Mhz 8x8
512GB Samsung SM951 M.2
Corsair 1500W 80 Plus Titanium AX1500i

On a 4K Monitor


Started logging on 04.04.16 19:08:30

OctaneRender 3.00 alpha 6 (3000006)

CUDA error 719 on device 1: An exception occurred on the device while executing a kernel. The context cannot be used anymore and must be destroyed. All existing device memory allocations from this context are invalid and must be reconstructed.
-> failed to launch kernel (ptLight1)
device 1: PMC kernel failed

[smicha]

Does this error also occur when using only 2nd gpu or with low priority set to on for the 1st gpu?

[Gineus]

It fails while SLI is on and off and with one or the other card active and on priority. I also used a try with just active and no priority.

I attached HW Monitor report (NOT active during other trials) of a fail with one card active (shows two have 84% mem usage ) and smbus report.

I also noticed that Octane crashes while a fail is loaded and I try to change the active card.

I hope this info helps

[Gineus]

I did some tests with 2.8 GB obj meshes and 2.2 GB meshes generated by Incendia 5 a totally differented process from above.

I generate them as STL and clean up in Meshlab (non manifold edges, Quadratic Edge reduction Laplace smoothing.

In above example meshes are generated by ZBrush and one by Sculptis and a basemesh from makehuman v1.

The fractal meshes are also not subdivided. The same result, There is a polylimit below graphics memory from where I am standing.

I used the same HDRI and on 4K and PMC


I used to fill up my 470 on Octane 1 and hit my polylimit on Octane 2.25 on 970 and on Titan X (SLI) quite often.

But I can definitely load more, so that's good.

I do a lot of fractal meshes as a hobby, so I am used to it.