SGI "Onyx4 UltimateVision" Uses ATI

Re: apache

It is. IIS 6 to be exact, running Windows 2003. We have no idea why our users are getting logged off either. We suspect the ads are doing something.

 

I would suggest that it is one GPU per pipe. Each pipe would be an independent viewport. Peak numbers SGI usually use are for combined pipes.

So peak scene complexity would be no different than expected from the consumer level boards (except of course that the overall system bandwidth is usually better).

btw (Can't remember who brought this up) the reality engine 2 graphics used to use Intel i860 cpu's for geometry calculations. (12)

 

Cheers. I knew the chips used were numbered as i?60. I believe it marked a change from their custom T&L acceleration to commodity parts but, then again, implementations tend to swing back and forth from custom to commodity to custom parts again.

 

Reality Engine used i860XP. (If you can find a copy of the paper "RealityEngine Graphics" by Kurt Akeley in the Siggraph 1993 proceedings, great read.)

DEC's PXG and PXG Turbo used a single i860 (and later the i860XP) as well, with a twist. The i860 was a geometry processor in early drivers on DECstations, but just functioned as not much more than a dma engine by the last drivers on AlphaStations.

-mr. bill

 

log off problem

Guys, try to uninstall the patch "q818529", this fixed the "unwanted log off" problem for me.

 

Yep. after RE2 they did go back to their own asics.

mrbill: Yes i860XP's 12 of them. 50Mhz with 100 single precision MFLOPS each. The i860XP also had rasterization hardware on board (32bit Gouraud/zbuffer no texture. actually programmable so first pixel shader unit? ). Obviously raster HW on i860XP was not used on the RE2.

 

More info on how pipes work together on UltimateVision.

Hi,

The Ultimate Vision machine uses the Origin 3xxx MIPS architecture, and connects many ATI cards as individual pipes. Current limit (by the look of it) is 16, but the aim is to go for more and more.

The pipes can either work independently or can be put through a network of compositors. At the moment, the compositors support spatial composition, so the NUMAflex architecture will be working hard to ensure that each pipe gets the right polygons.

In the future, Z and A composition should be possible, which will make parallelism much easier to achieve.

I ran tenmillion (which tries to get the maximum polys/sec), and I seem to recall that I easily got 80M polys/sec on it (don't quote me! this was about 2 months ago), but what was especially pleasing was that I got 160M polys/sec going to two pipes.

Just good to see that theoretical number being reached. Of course for real parallelism, your mileage might vary!

SchoolBully