PDA

View Full Version : SGI Onyx4 UltimateVision Comments?



Husted
07-16-2003, 12:07 AM
Hi,

At the risk of being moderated, I hereby post a slightly off-topic question:

Anyone care to comment on the new SGI Onyx4 UltimateVision?

Product info, see:

http://www.sgi.com/vizuliation/onyx4/

Best Regards,


Niels

dorbie
07-16-2003, 08:57 AM
Summary, ATI chips each on a PCI bus with back end digital video level compositing of some sort and multithreaded application software or middleware on the front end required to feed it.

Husted
07-16-2003, 10:45 AM
Hi,

The information that I dug up states that the Onyx4 UltimateVision is equipped with a number of ATI FireGL X1 boards (R300) - but I have not been able to find much information on the compositor.

So far it seems as if there is no hardware support for compositing - yet. So if you want to exploit multiple pipes, you need to do this manually using the OpenGL Multipipe SDK.

The Scalable Graphics Compositor - when it arrives - seems to be using DVI signals as input.

So if I get this right, you need to encode the depth buffer in the frame buffer of the 2nd channel to perform depth compositing?

How can the compositor be used to composite floating point buffers if it use DVI inputs?

-- Niels

dorbie
07-16-2003, 11:56 AM
There are other ways to do the composition with no depth. You spatially partition the data and sort & switch the channel composition by depth. You do need alpha out though. As a bonus you get full antialiasing support. There are other simpler approaches too, it depends on what you want to scale, and how well.

Hmmm, so no compositor yet, that bites. It's the most interesting piece of the puzzle and you are telling me they don't have it. I thought they'd already done this for their vpro based system, what a shame.

Husted
07-16-2003, 12:24 PM
Hi,

It seems that the depth compositing today is supported through the monster mode in the OpenGL Multipipe SDK - but essentially, this means reading the pixels using glReadPixels, and doing a software depth compositing. http://www.opengl.org/discussion_boards/ubb/frown.gif

I also suspect that they use a PCI version of the FireGL boards (Are there any AGP->PCI bride out there?). How is the pixel readback over the PCI bus?

I was hoping that they found a nice way to do depth compositing - without reading back the color and depth buffer.

-- Niels

Ysaneya
07-16-2003, 12:28 PM
I can't say about the technical details, but i know for sure that read backs are extremely fast on the Onyx; reading back the ZBuffer can be done in real-time.

I think they're not using a standard PCI bus, but something called "PCI express". Google it for more infos.

Y.

TriangleMan
07-16-2003, 12:43 PM
Angus is right, the compositing is the interesting piece, that would make this competitive.
I have not heard of actuall hardware working, so it is what is termed as vaporware now, right?

The other interesting piece to this is the syncing solution across multiple pipes. The Onyx does this (in full genlock) marvolously. What does SGI plan to use here? It would be really nice if someone from SGI would state what it is, and if it actually exists, and actually works.
Works being defined as running with more than two pipes at once, with vertical blanking synchronized, and not having completely insane rebooting procedures, or kernel drivers that hog the CPU.

If this exists, I would be very interested.

Husted
07-16-2003, 01:24 PM
Hi,


Originally posted by TriangleMan:

...
I have not heard of actuall hardware working, so it is what is termed as vaporware now, right?

The other interesting piece to this is the syncing solution across multiple pipes. The Onyx does this (in full genlock) marvolously. What does SGI plan to use here?
...


There is already compositing bricks for the Onyx3, I suspect it is the same bricks they use for image space compositing right now. See: www.sgi.com/visualization/onyx/3000/ip/bricks.html (http://www.sgi.com/visualization/onyx/3000/ip/bricks.html)

for more details on the current compositing brick.

Regarding synchronization, they use their ImageSync technology (Also used in their former Linux Graphics Cluster). Basically this is synchronization based on a number of interconnected serial boards.

On the graphics cluster, the ImageSync board was connected to the graphics board (NV20/NV25 based boards). The ImageSync technology ensured that the drift between the boards was less than a single scan line.

The only problem I see with the ImageSync technology is that you cannot connect to an external genlock signal.

-- Niels

[This message has been edited by Niels Husted Kjaer (edited 07-16-2003).]

dorbie
07-16-2003, 07:00 PM
Readback to the host is not very fast on an original Onyx, I've done some work in this area, readback of color and Z would be slow and defeat antialiasing for compositing, But this is completely different anyway, I don't think they use PCI express yet, infact I'm pretty sure they don't.

Scalability depends on your target frame rate and data size, it's certainly possible for big databases and modest frame rates. Readback with alpha might make more sense, it depends on the structure of your software. AA rendering aginst black gives you a classic premultiplied AA scene sections ready for sorted composition.

The digital compositing capability on some earlier Onyx systems depends on very specialized display generator boards, other systems would require external switching video compositors but required specific digital video output from graphics.

It's not wise to simply assume all this is supported on this new system.

m2
07-17-2003, 12:47 AM
Originally posted by dorbie:
ATI chips each on a PCI bus with back end digital video level compositing of some sort and multithreaded application software or middleware on the front end required to feed it.

I'm blind, where did you read that the system has ATI chips? I know the information is correct, I just can't find it on the website. The closest I found was "off the shelf components".

This reminds me of a quote from a HP engineer regarding their new EV7-based systems which sport an ATI card and run on OpenVMS: "NVIDEA (sic) WILL NOT GIVE YOU THE TIME OF DAY, LET ALONE SPECS, CODE OR HELP. ATI WILL AT LEAST GIVE YOU THE TIME OF DAY." Perhaps NVIDIA doesn't like it when people can't spell its name :-)

Husted
07-17-2003, 01:02 AM
Hi,


Originally posted by m2:
I'm blind, where did you read that the system has ATI chips? I know the information is correct, I just can't find it on the website...

You'll find this information on news.com (Found the link from beyond3d.com)
http://news.com.com/2100-1010_3-1025324.html

-- Niels