As far as I remember from watching the specs in order to high a better rate
the magic number is the memory speed transfer.
I was going to buy a 680 but then realized that my 470 has more speed in
the memory transfer.
Don't know why but Nvidia choked the speed in most of the 600 series except
for the Titan. And open again the hoose in the 690.
In my experience and from what we have discussing in the Redshift forums
the two magic numbers to see which GPU will perform faster is not only the
Fill triangles per sec number but also the memory transfer speed. Below
are the specs of the Titan, the 680 and the 470 I have.
*GTX TITAN GPU Engine Specs:*
2688CUDA Cores
837Base Clock (MHz)
876Boost Clock (MHz)
187.5Texture Fill Rate (billion/sec)
*GTX TITAN Memory Specs:*
6.0 GbpsMemory Clock
6144 MBStandard Memory Config
GDDR5Memory Interface
384-bit GDDR5Memory Interface Width
288.4Memory Bandwidth (GB/sec)
*GTX 680 GPU Engine Specs:*
CUDA Cores
1006Base Clock (MHz)
1058Boost Clock (MHz)
128.8Texture Fill Rate (billion/sec)
*GTX 680 Memory Specs:*
6.0 GbpsMemory Speed
2048MBStandard Memory Config
256-bit GDDR5Memory Interface Width
192.2Memory Bandwidth (GB/sec)
GTX 470 GPU Engine Specs
CUDA Cores 448
Graphics Clock (MHz) 607 MHz
Processor Clock (MHz) 1215 MHz
Texture Fill Rate (billion/sec) 34.0
Memory Specs
Memory Clock 1674 MHz (3348 data rate)
Standard Memory Config 1280 MB
Memory Interface GDDR5
Memory Interface Width 320-bit
Memory Bandwidth (GB/sec) 133.9 GB/sec
So as you can see my 470 MB (Memory Bandwith) is not that far from the 680
but the Memory Clock speed, and the Processor Clock speed, of the 470 is
higher than the 680
Not a guru here but that can be an explanation of why the 470 has faster
fps than the 680...
For rendering you make sure you compare the Memory Bandwith of the GPU.
Beside the CUDA. Double the CUDAs but half the Memory Bandwith, and I
cannot assure but it will render almost the same.
2014/1/9 Leonard Koch <leonardkoch...@gmail.com>
> I get about 28-31 out of my 680. Does anyone have a common explanation for
> that?
>
>
> On Thu, Jan 9, 2014 at 5:10 PM, Emilio Hernandez <emi...@e-roja.com>wrote:
>
>> Hey Mirko I ran your script and I got 50.7 fps...
>>
>> But then I remembered I have my displays plugged in to my 470.. hahaha.
>>
>> Don't ask why, but when using AE with the displays plugged into the Ti,
>> AE does not like it and disables GPU for calculations...
>>
>> Pffff.
>>
>>
>>
>>
>> 2014/1/9 Mirko Jankovic <mirkoj.anima...@gmail.com>
>>
>>> Hey Tim
>>> Would you be able to take 2 minutes of your tmie and run this ol python
>>> script for SI with your titan?
>>> I'm getting weird results with an 780 in my home system outperforming
>>> titan a lot... well here is copy paste from forum if you are able to check
>>> it out as well.. thanks!:
>>>
>>> itan: ~170 fps
>>> 780: ~245 fps
>>>
>>> Go figure [image: :)]
>>> But I'm suspecting something weird with my titan system for some time
>>> will have to test further but would be great if anyone with titan as well
>>> could run it too?
>>> This old python script:
>>> Application.CreatePrim("Cube", "MeshSurface", "", "")
>>> Application.SetValue("cube.polymsh.geom.subdivu", 831, "")
>>> Application.SetValue("cube.polymsh.geom.subdivv", 800, "")
>>> Application.SetValue("cube.polymsh.geom.subdivbase", 800, "")
>>> Application.SetValue("Camera.camvis.refreshrate", True, "")
>>> Application.SetDisplayMode("Camera", "shaded")
>>> Application.DeselectAll()
>>> Application.SetValue("PlayControl.Out", 5000, "")
>>> Application.DeselectAll()
>>> Application.GetPrim("Null", "", "", "")
>>> Application.SelectObj("Camera_Root", "", "")
>>> Application.CopyPaste("Camera_Root", "", "null", 1)
>>> Application.SelectObj("null", "", "")
>>> Application.SaveKey("null.kine.local.rotx,null.kine.local.roty,null.kine.local.rotz",
>>> 1, "", "", "", "", "")
>>> Application.SetValue("PlayControl.Key", 5000, "")
>>> Application.SetValue("PlayControl.Current", 5000, "")
>>> Application.Rotate("", 0, 8000, 0, "siAbsolute", "siPivot", "siObj",
>>> "siY", "", "", "", "", "", "", "", 0, "")
>>> Application.SaveKey("null.kine.local.rotx,null.kine.local.roty,null.kine.local.rotz",
>>> 5000, "", "", "", "", "")
>>> Application.FirstFrame()
>>>
>>> Just paste in python script run and hit play.
>>> Thakns!
>>>
>>>
>>> On Thu, Jan 9, 2014 at 3:34 PM, Tim Crowson <
>>> tim.crow...@magneticdreams.com> wrote:
>>>
>>>> We've been testing 1 Titan vs. 3 and so far, the speed increase of the
>>>> triple-Titan box is holding at about 2.45x. In an email exchange (or maybe
>>>> it was on the forums, can't recall) it was mentioned that on the topic
>>>> parallelization, Pixar had determined that even for them, 4 units together
>>>> (of whatever, not necessarily Titans) was the max they could really go
>>>> before it started to cost more money than it was worth. In our case, I'm
>>>> thinking 3 might be our max, based on some nerdy mathematics by one of our
>>>> IT guys analyzing render times per shot, per frame, hardware/software
>>>> costs, rack space used, etc.
>>>>
>>>> But hey, Redshift aside, the Titan in my workstation is doing wonders
>>>> for my viewport performance in Soft. I had a 58M, 2500-item model derived
>>>> from a CAD file the other day, and this thing was letting me tumble around
>>>> it at ~15fps in Shaded mode. That ain't shabby!
>>>> -Tim
>>>>
>>>>
>>>>
>>>> On 1/9/2014 6:11 AM, Paul Griswold wrote:
>>>>
>>>> There was a discussion on the RS forums about it. I don't recall the
>>>> numbers, though. I don't think the speed of the PCIe slot made a huge
>>>> difference. It's really all about the speed of the card.
>>>>
>>>> Also, although it doesn't load the entire scene into your card's
>>>> memory, the more memory your card has, the better it is.
>>>>
>>>> But overall, for the type of work I'm mainly doing these days, it's
>>>> extremely fast. In fact, it's so fast that I was finding the bottleneck
>>>> was the time taken to export the mesh to Redshift, not rendering. Redshift
>>>> has a proxy system like Vray & Arnold, but you have to manually create
>>>> proxies per object & my scene had hundreds and hundreds of objects, so I
>>>> didn't have time to create them. Therefore, it was creating a renderable
>>>> mesh per frame - so on a frame that took 28 seconds to render, 20 seconds
>>>> was spent exporting the mesh and 8 seconds were spent on rendering. But
>>>> again, it's a beta and they're continuing to improve things like the proxy
>>>> system.
>>>>
>>>> Once I'm caught up I'm hoping to try rendering the classroom scene
>>>> and see how it does.
>>>>
>>>> -Paul
>>>>
>>>>
>>>> ᐧ
>>>>
>>>>
>>>> --
>>>>
>>>>
>>>>
>>>
>>
>
