Welcome to our community

Be a part of something great, join today!

Log in
  • Hey all, just changed over the backend after 15 years I figured time to give it a bit of an update, its probably gonna be a bit weird for most of you and i am sure there is a few bugs to work out but it should kinda work the same as before... hopefully :)

BOXX rocks Redcode on Intel Skull Trail!!

luki said:
Guys,

It is a 4K R3D file played with the 1/2 Rez High setting. It is pulling from the 2K layer, not the 4K layer in the R3D file.

But it still a huge achievement...

Lucas
-----
ASSIMILATE, Inc.
LA, CA, USA
Click to expand...

Its a fantastic achievement indeed but my 8 core MacPro can do it too. So its nothing specific to the Boxx box.
 
Emery said:
Its a fantastic achievement indeed but my 8 core MacPro can do it too. So its nothing specific to the Boxx box.
Click to expand...

But your 8 core MacPro can't get HDSDI out from the GPU.

The significance here is that RIGHT NOW we can got to SR wit a full 2K debayer. That is significant in my book for sure.

And I sincerely doubt your Mac it will do a 20 minute reel sustained at 24 fps - but maybe it can - I'll check that.
 
Offhollywood said:
But your 8 core MacPro can't get HDSDI out from the GPU.

The significance here is that RIGHT NOW we can got to SR wit a full 2K debayer. That is significant in my book for sure.

And I sincerely doubt your Mac it will do a 20 minute reel sustained at 24 fps - but maybe it can - I'll check that.
Click to expand...

Yea no HDSDI love on the Mac. But even on a BOXX we couldn't get that for RC, just Scratch. Exciting all the same.
 
Lexicon said:
Jeff, if you bothered to read what I said earlier, I said it makes no sense for them to have listed the board's supported amount of memory at 8GB in the online specs they have because we all know the 5400 chipset is for workstations and it is obviously not limited at 8GB
Click to expand...

Fair enough... I didn't interpret it that way. I read it as you saying it didn't make sense that they would cap the RAM at 8GB, not that the spec page was in error.
 
Emery said:
Yea no HDSDI love on the Mac. But even on a BOXX we couldn't get that for RC, just Scratch. Exciting all the same.
Click to expand...

If somebody would try out Nvidia Quadro 5600 HDSI card with Mac Pro 8 core to see if it could work with Booth Camp Windows XP...then maybe you get HDSDI out of Mac...
 
I am wondering about one thing:
1/2 rez is 2k, right?
But 2k would be quarter rez of 4K? Is it referring to the horizontal resolution only?

Skulltrail is nice for an midprice system, however if you really want topclass performance, there is no way around even more powerful systems as dvs clipster or hyrdas - these play multistream 4k, uncompressed, with realtime fx and colorcorrections.
Another option are codecs as cineform, which also do 4k rt.

Two question reagarding SLI on the skulltrail -
- can scratch already use multiple graphic cards (for decoding 2 images at the same time, for 3D or higher framerates in higher resolutions)?
- where is the bottleneck, is it cpu oder gpu? are all 8 cpus loaded to max or?

The speedgrade system had it bottleneck in the nvidia gpu (was only 4500/4600 w/o sli however).
 
laguun said:
I am wondering about one thing:
1/2 rez is 2k, right?
But 2k would be quarter rez of 4K? Is it referring to the horizontal resolution only?
.................
- where is the bottleneck, is it cpu oder gpu? are all 8 cpus loaded to max or?
Click to expand...

I would say yes, yes and probably yes! :biggrin:
Re bottleneck I believe its CPU...
 
laguun said:
I am wondering about one thing:
1/2 rez is 2k, right?
But 2k would be quarter rez of 4K? Is it referring to the horizontal resolution only?

Skulltrail is nice for an midprice system, however if you really want topclass performance, there is no way around even more powerful systems as dvs clipster or hyrdas - these play multistream 4k, uncompressed, with realtime fx and colorcorrections.
Another option are codecs as cineform, which also do 4k rt.

Two question reagarding SLI on the skulltrail -
- can scratch already use multiple graphic cards (for decoding 2 images at the same time, for 3D or higher framerates in higher resolutions)?
- where is the bottleneck, is it cpu oder gpu? are all 8 cpus loaded to max or?

The speedgrade system had it bottleneck in the nvidia gpu (was only 4500/4600 w/o sli however).
Click to expand...

Hi Laguun...

Yes, 2K is 1/2 the resolution of 4K. Yes, Full Rez realtime decode would be better. But the 1/2 Rez high looks amazing... we are currently doing tests at several SCRATCH customers and letting them determine how good it is for their needs. I'll report more when I know more.

The question about multiple graphics cards should be addressed more to ATI and NVidia than ASSIMILATE. There currently just isn't a very good way to tie GPUs together for high framerate, high-bandwidth texture transfer. SLI is built for gaming and 3D, not DI. The Quadroplex architecture also doesn't work for DI, so that's not an option. And Tesla uses the same underlying principles as Quadroplex. So until some fairly basic things change, we have to stick to one.

That's not currently such a big issue, though. The 4600 generation doubles the texture memory. This allows for a dramatic increase in realtime capability... up to and including 4K RGB.

Lucas
------
ASSIMILATE, Inc.
LA, CA, USA
 
luki said:
Hi Laguun...

The question about multiple graphics cards should be addressed more to ATI and NVidia than ASSIMILATE. There currently just isn't a very good way to tie GPUs together for high framerate, high-bandwidth texture transfer. SLI is built for gaming and 3D, not DI.
Click to expand...
I wouldn´t say so.
The drivers allow to transfer individual frames through individual bus architectures to individual processing pipelines with driver support. i see plenty of new improvements for DI ahead, stereoscopic display being certainly one of the first.

we get -huge- performance boosts with VFX systems in SLI configurations, nowhere near 100%, but easily 40-80%, especially in higher resolutions as 2560*1600, where a single chips memory interface is the bottleneck on 5400er systems (aka skulltrail).

depending on chipset and driver, you get indeed double transfer bandwidth with sli.

luki said:
The Quadroplex architecture also doesn't work for DI, so that's not an option. And Tesla uses the same underlying principles as Quadroplex. So until some fairly basic things change, we have to stick to one.
Click to expand...
Yeah, the busses as limitation... Tesla & Quadro are no real advance for transfer lmited applications.

luki said:
That's not currently such a big issue, though. The 4600 generation doubles the texture memory. This allows for a dramatic increase in realtime capability... up to and including 4K RGB.
Click to expand...
Yeah, the G80 chips as the geforce 8800gts (aka Quadro 4600) and 8800GTS (aka Quadro 5600) have doubled the memory bandwidth as well, and increased texture buffers to 1024 on the 8800gts and 1536 on the 8800 gtx.

However, i am curious, how do you expect to quadruple the performance from 2K RT to 4K RT (which is 4 times the res) by having the doubled memory size and bandwidth in nvidias chips? lower quality debayer then?
 
Doesn't SLI tie two cards together and the Master card encodes even frames while the slave does odd frames in Alternate Frame Rendering, AFR? Then the data is sent to the Master card over the SLI bridge for output, doubling core clock, Memory, Frame Rate and Bandwidth? Over one output? Or is the DI Hard/software not taking advantage of these factors or can it not take advantage? I'm looking for more of an explanation I guess.
 
I think SLI is optimized for games. In games, you load a bunch of textures into graphics card memory mostly when you start a level, etc. Once everything is loaded, the bottleneck is GPU speed combined with speed of transfer from graphics card memory to GPU.

In DI, you're constantly loading textures from main memory onto the GPU, right? This is a different operation which presumably hasn't been optimized for SLI yet.

Bruce Allen
www.boacinema.com
 
luki said:
Good idea, Sanjin. You're nominated. Let us know your results. :)

Lucas
-----
ASSIMILATE, Inc.
LA, CA, USA
Click to expand...

Luki,

for sure I'll do it if somebody would bring me the card (free) to test :) .
 
We just got a new MacPro ordered, we got a new Nvidia 4600 yesterday and we own and run a Scratch here in Paris.

So if by any chance we can get our new MacPro to accept the 4600 we could try to run Scratch under bootcamp and compare results with the ones Mark got.

By the way Mark, great results, we were also working on the skulltrail route for speeding up our Scrath PC side, but you definitely got much ahead of us on this one! We were still struggling on getting prototype 800 FB-DIM from Kingstons, but we were told that there was none available due to an 'intense collaborative project with BOXX'... now I see who was behind all this!

Lucas, I was with Daniel and Nacho doing a Red/Scratch demo in Paris yesterday and I was thinking about talking to Alain from SIMMS about this Skulltrail option. I think we should maybe bring that discussion to the Scratch usergroup as well, or post a link towards this thread if it's ok with you.
 
Bruce Allen said:
In DI, you're constantly loading textures from main memory onto the GPU, right? This is a different operation which presumably hasn't been optimized for SLI yet.

Bruce Allen
www.boacinema.com
Click to expand...

I believe it really just depends on how you want to go about it and how beefy the card is when it comes to GPU and texture memory. We've certainly got cards with 1GB of memory on them these days. Using multiple GPUs to process a single frame of RED footage could be quite hairy (but I think RED might pull it off if given the proper support from Nvidia and ATI). The ideal and probably easiest method to get working with an SLI configuration is to simply use AFR (Alternate Frame Rendering) which basically renders one frame at a time on each GPU. Quad SLI would give you four frames at once in the pipeline. Granted you would likely want to have something close to a very powerful server for it but I think that could be likely done without any human sacrifices on behalf of the RED gods. :biggrin:
 
Arno said:
Lucas, I was with Daniel and Nacho doing a Red/Scratch demo in Paris yesterday and I was thinking about talking to Alain from SIMMS about this Skulltrail option. I think we should maybe bring that discussion to the Scratch usergroup as well, or post a link towards this thread if it's ok with you.
Click to expand...

Totally cool with me... post away. : )

Lucas
 
I agree, GPU bandwidth is a huge concern. NVIDIA claims up to 2x the performance of one card with SLI enabled. Using AFR, I believe Redcine at least could take heavy advantage of this, One would think with a 3 GB/s internal Hardware Raid and SLI, 2k High Redcode would be more achievable on slower machines (4-core). I doubt the real bottleneck is with the CPU.

The Skull Trail is nice but its obviously lacking in some memory depts... I'm happy to see good OC features, and Quad SLI on a workstation board though... Some dont have any OC at all...!
 
Lexicon said:
The ideal and probably easiest method to get working with an SLI configuration is to simply use AFR (Alternate Frame Rendering) which basically renders one frame at a time on each GPU. Quad SLI would give you four frames at once in the pipeline.
Click to expand...

Yes... if you could keep the GPUs fed with data. I'm saying that nVidia probably didn't have constantly sending different 4k textures to each GPU in mind when they designed and tested their AFR implementation. It's probably good at rendering 4 frames at once with the same texture info, not with different texture info. So the GPUs are probably spinning their wheels in a confused fashion wondering what the heck is going on half of the time.

The easiest way to fix this is to invent a RedCineBench and get all of the hardware review sites to include this in their testing - it's well known that the graphics card manufacturers optimize for these ;)

BTW, Sanjin, I would be prepared to bet money that a new 8-core Mac running Windows in Boot Camp mode would work very well indeed with a Quadro SDI.

Also, how much is Skulltrail? Some of the prices guessed at on Tomshardware etc seemed ludicrously high - eg the same as just getting an 8-core Mac Pro. If it were at all close, why not get a Mac plus Bootcamp? To me, the amazing value in PCs is in the midrange.

Bruce Allen
www.boacinema.com
 
luki said:
With decompressing and debayering R3D material, the bottleneck is indeed the CPU.

Lucas
-----
ASSIMILATE, Inc.
LA, CA, USA
Click to expand...

Do you see debayer computation moving from cpu to gpus in the G80 architecture?
 
You must log in or register to reply here.
Back
Top