Multiple R300

[OpenGL guy]

About simFUSION 6000

simFUSION 6000 offers unbeatable PC-IG performance and image quality, with 2-to-24-sample full-scene antialiasing at screen resolutions up to 2048x1536 and fill rates up to 9.6Gpix/sec

I hope they stuffed up that fill rate because that is 1/8th of what I'd hope for a 4 GPU solution.

I don't understand your comment at all. 9.6 Gpix/sec = 4*2.4Gpix/sec = 4*8*300 which implies the R300s are clocked at 300 mhz (I don't know if this is true for a fact, just basing it off the released numbers).

Why would you expect four chips to get more than 4x the actual fillrate of a single chip?

To get the expected ~ 80 GB/sec bandwidth wouldn't each GPU need the fully 128MB of memory, implying a 512MB card == very expensive.

I don't think price is a consideration here.

I just smile that this innovation that served 3dfx so well has been re-introduced not by the company that bought them but using their arch rival's technology...

This isn't the same method as 3dfx used...

 

[noko]

This is very interesting technology and it actually looks like it can be useful in doing real time rendering with the advance Pixel shaders of the R300. Much more powerful by far then a single NV30 design cinematic board. I think ATI is very much on the right track here to get into professional video production and special effects department of many studios. ATI continues to let out surprises after surprises with bold partners making products that work at the forefront of technology.

Now come to think of it, this setup could do some pretty neat real time imaging for like medical stuff. Hmmm maybe another market, I really don't know but it is a thought.

 

[g__day]

OpenGL guy / or others

Have I mis-interpreted the stated 9.6 Gigapixel / second fill rate as the memory bandwidth (the Radeon 9700 Pro has a memory bandwidth of about 19.6 GB/sec).

I had about 4 hours interrupted sleep last night (kids are my joy). Did I simply mistake the gigapixel fill rate (is that the GPUs combined shader theoretical performance?) for the memory bandwidth figure?

Apolpgies if I confused matters.

 

[Althornin]

OpenGL guy / or others

Have I mis-interpreted the stated 9.6 Gigapixel / second fill rate as the memory bandwidth (the Radeon 9700 Pro has a memory bandwidth of about 19.6 GB/sec).

I had about 4 hours interrupted sleep last night (kids are my joy). Did I simply mistake the gigapixel fill rate (is that the GPUs combined shader theoretical performance?) for the memory bandwidth figure?

Apolpgies if I confused matters.

yeah you did.
i was like WTF?
but its cool. We wont break your kneecaps this time

 

[Mulciber]

This is very interesting technology and it actually looks like it can be useful in doing real time rendering with the advance Pixel shaders of the R300. Much more powerful by far then a single NV30 design cinematic board. I think ATI is very much on the right track here to get into professional video production and special effects department of many studios. ATI continues to let out surprises after surprises with bold partners making products that work at the forefront of technology.

Now come to think of it, this setup could do some pretty neat real time imaging for like medical stuff. Hmmm maybe another market, I really don't know but it is a thought.

Im still wondering why nVidia and Quantum 3d never released a 4 or 8 way nv25. I can only hope that they will with the nv30, and watercooled, otherwise it will be an absolute joke. I think the nv30 would be a much more likely candidate for an offline renderer considering its 128bit frame buffer support.

I still can't believe that neither ATI nor nVidia have done a dual chip board in their FireGL or Quadro lineups (not talking about multi-monitor support quadros).

 

[KimB]

Im still wondering why nVidia and Quantum 3d never released a 4 or 8 way nv25. I can only hope that they will with the nv30, and watercooled, otherwise it will be an absolute joke. I think the nv30 would be a much more likely candidate for an offline renderer considering its 128bit frame buffer support.

I still can't believe that neither ATI nor nVidia have done a dual chip board in their FireGL or Quadro lineups (not talking about multi-monitor support quadros).

They probably didn't see any interest. I think that only recently with the very highly programmable nature of these new processors do they think they can actually find somebody to sell these things to.

I fully expect most high-end movie-type 3D rendering work to be done on similar graphics cards within three years.

 

[T2k]

Now come to think of it, this setup could do some pretty neat real time imaging for like medical stuff. Hmmm maybe another market, I really don't know but it is a thought.

Neither display nor videocard is the problem for the medical illustration/imaging/animation. Since there almost everything based on volume graphics, only the memory limits your work. That's the main reason why every company combines the volume renders with the polygon-based techniques.

 

[Tagrineth]

Im still wondering why nVidia and Quantum 3d never released a 4 or 8 way nv25.

How would the chips communicate? Nothing before NV30 has the capacity for properly functional multichip. Sure they can use screen subdivision on a driver level but that's incredibly inefficient...

 

[Dave Baumann]

How would the chips communicate? Nothing before NV30 has the capacity for properly functional multichip

There are ways it can be forced outside of the chip. Quantum have done it with V3's and other chips despite having no in-built functionality.

Besides its been said frequently that NVIDIA has had AGP bridges built into their products for multichip capabilities since GF2.

 

[Dio]

How would the chips communicate? Sure they can use screen subdivision on a driver level but that's incredibly inefficient...

Scissor clip registers make screen subdivision trivial.

Resolving the output is the tricky bit, but there are all sorts of ways to get around that.

 

[T2k]

Im still wondering why nVidia and Quantum 3d never released a 4 or 8 way nv25.

How would the chips communicate? Nothing before NV30 has the capacity for properly functional multichip. Sure they can use screen subdivision on a driver level but that's incredibly inefficient...

Rage Fury MAXX

 

[Tagrineth]

Im still wondering why nVidia and Quantum 3d never released a 4 or 8 way nv25.

How would the chips communicate? Nothing before NV30 has the capacity for properly functional multichip. Sure they can use screen subdivision on a driver level but that's incredibly inefficient...

Rage Fury MAXX

nVidia, not ATi. ;P

And with screen subdivision like what's being suggested, wouldn't the performance be a little... odd? I seem to remember a test sample of a dual GTS which had huge problems involving one chip stalling constantly due to its render load being far less than the other's.

 

[Dio]

There are pathological cases for ALL architectures - because some chips own some pixels and some others. I wouldn't expect real world applications to have much problem even with just the scissor method.

 

[arjan de lumens]

With screen subdivision one could conceivably use an adaptive algorithm adjusting the exact subdivision based on the actual workload of each GPU, giving each GPU less of the screen to work with if it is overworked and more if it has too little to do per frame. This might give a bit uneven framerates if the scene changes too dramatically from one frame to the next, though.

 

[Panajev2001a]

As mentioned, bandwidth does scale: Each chip has it's own dedicated 256-bit bus. The "unified" comment that was made was referring to the fact that it's a unified bus for texture, depth and color data, just like pretty much every chip made after the Voodoo 2.

cough... Graphics Synthesizer... cough...

 

[OpenGL guy]

cough... Graphics Synthesizer... cough...

I did say "pretty much"

 

[T2k]

Do you remember CAE and their 4-chips solution, called TROPOS, based on R200s?

They've launched the Medallion-S, in 4/8/12/16 GPUs configurations -based on R300-version...

Specs are here: http://www.cae.com/en/visuals/pdf/medallion_s_specs.pdf

 

[Doomtrooper]

Yes..and lets not forget the Dual R100 board that was on sale on Ebay..engineering sample but it did exist...too bad the pics don't work anymore. This was legit, no photoshop job.

http://www.rage3d.com/board/showthread.php?s=&threadid=33626393&highlight=multiple+chip

 

[sireric]

With screen subdivision one could conceivably use an adaptive algorithm adjusting the exact subdivision based on the actual workload of each GPU, giving each GPU less of the screen to work with if it is overworked and more if it has too little to do per frame. This might give a bit uneven framerates if the scene changes too dramatically from one frame to the next, though.

Interesting. However, I should point out that assuming that the screen is divided into 4 tiles for four chips is wrong for the R300 based E&S solution. The tile size, while not revealed is much smaller than 1/4 screen, so that there's many tiles for each chip, distributed. Load balancing is nearly perfect in most cases (3dmark, for example, is linear with # chips), though I'm sure that something like a single pixel poly case could show limits. But since the target is a simulation environment at high resolution, the balancing should be very even.

An adaptive tile size would be interesting, but I can't see that being more dynamic than once per frame. Of course, it really would only work if you had one or a few tiles per chip, otherwise trying to figure out a re-tiling of the screen would be non-trivial. I guess you could have different configurations and select betwee them, but even there, I only see it working easily with one tile per chip.

Later