OTOY Forums

Question for 100 points - why the scenes in OCtan load to the GPU memory so long and what can speed up this process?

colin wrote:@jaberwocky: as i understand there is no possibility to combine an ssd raid with a conventional hdd in the same system, as of yet. raid hdd + ssd works just fine. raid ssd on its own as well. but that wouldn't be a practical set up i think.
plus i believe there's a 750 MB/s speed limit. but don't quote me on that as i actually don't know much about that topic.

The 750MB/s PCIe SSD´s use a raid system.
You can also buy a PCIe SSD 1 TB from OCZ with 1500MB/s for 3.879€

face

FamilyGuy wrote:Question for 100 points - why the scenes in OCtan load to the GPU memory so long and what can speed up this process?

If I am correct ,The Scene is currently processed (Voxalised) using one thread of a CPU and stored using the motherboards memory.The file is then sent to the graphics cards memory via the PCIE interface and stored there where the G.P.U. then access's it to render the scene.Next frame, the same process happens all over again.The only way to speed this up easily I guess is to write that part of Octane that Voxalises the scene to take advantage of the current quad core or 6x core CPU's available.Thus Speeding up the voxalising part of the process by 4-6 times.


Errr..... do i get my 100 points now

FamilyGuy wrote:Question for 100 points - why the scenes in OCtan load to the GPU memory so long and what can speed up this process?

I think that is because they read in an ascii file and change it to real values.
Thats also the problem for exporting.
I has tested it and came not on the write speed of a HD, so that a saving to a ram/ssd can´t speedup it dramatically.
The only thing to speedup it, is to read the values from an app and write it direct in the GPU, also DAI (deep application integration)
But for this and the mass of supported apps, it should take some month, to implement it...

face

Jaberwocky wrote:

FamilyGuy wrote:Question for 100 points - why the scenes in OCtan load to the GPU memory so long and what can speed up this process?

If I am correct ,The Scene is currently processed (Voxalised) using one thread of a CPU and stored using the motherboards memory.The file is then sent to the graphics cards memory via the PCIE interface and stored there where the G.P.U. then access's it to render the scene.Next frame, the same process happens all over again.The only way to speed this up easily I guess is to write that part of Octane that Voxalises the scene to take advantage of the current quad core or 6x core CPU's available.Thus Speeding up the voxalising part of the process by 4-6 times.


Errr..... do i get my 100 points now

The problem is, that you can only read/write sequentially from a file.
If you will write with more cores to the GPU, you must know the position of the previous value, otherwise you overwrite they.
You can solve it with an index on which evere core have access and which the process (core) can set a status that the value is written.
But i think, the whole expense cost some time and you will have on a quadcore max a speedup of 2.5-3.
Its also not so simple to implement it...

face

Edit:
Ok, you can read/write multiple times from a file, but it coast realy much performance...

face wrote:

Jaberwocky wrote:

FamilyGuy wrote:Question for 100 points - why the scenes in OCtan load to the GPU memory so long and what can speed up this process?

If I am correct ,The Scene is currently processed (Voxalised) using one thread of a CPU and stored using the motherboards memory.The file is then sent to the graphics cards memory via the PCIE interface and stored there where the G.P.U. then access's it to render the scene.Next frame, the same process happens all over again.The only way to speed this up easily I guess is to write that part of Octane that Voxalises the scene to take advantage of the current quad core or 6x core CPU's available.Thus Speeding up the voxalising part of the process by 4-6 times.


Errr..... do i get my 100 points now

The problem is, that you can only read/write sequentially from a file.
If you will write with more cores to the GPU, you must know the position of the previous value, otherwise you overwrite they.
You can solve it with an index on which evere core have access and which the process (core) can set a status that the value is written.
But i think, the whole expense cost some time and you will have on a quadcore max a speedup of 2.5-3.
Its also not so simple to implement it...

face

Edit:
Ok, you can read/write multiple times from a file, but it coast realy much performance...

I'm not really an expert.But maybe you could comment on if they could use this to optimise their code.


see link

http://software.intel.com/en-us/article ... rallel_(S)

So to put it in short words:

SSDs are great but they won't speed up Octane's Mesh import / export dramatically.