Multiple R300

[Althornin]

Unified.

Heh.
I'd trust sireric over PR-speak any day.

 

[jvd]

It would be nice to see a dual r350 out. even if it does cost 800 bucks ati could just make a limited run of it say mabye a couple thousand boards and sell them through thier site. I couldn't afford it but it would be nice to see out there and it wouldn't be a bad investment. Most likely the r400 and nv40 wont surpass the speed in current games. Hell if it could play star wars galaxis with everything maxed at 1600 x 1200 with say 6fsaa and 16 aniso I'd actually save and buy one... But it wont happen. Mabye in a few generations we will see multi chip cards do to the sheer size of the chips

 

[Saem]

One could say the V5500 had a non-unified 256bit bus. This could be a jab at a Quantum3D product which has a non-unified 256bit bus.

 

[Vince]

Unified.

Or would this be to differentiate between solutions that have devoted resources strictly for the framebuffer and textures.

 

[Chris]

I to am very interested in how geometry is handled. Adding parallelism to the processing of geometry across the multiple chips is enticing but it seems as though the subdivision of scene geometry would have to happen at the driver level.

Could all four chips be ‘listening’ so that scene geometry is sent only once? Each chip receives full scene geometry but only renders the viewspace for which it is responsible?

Regards, Chris.

 

[Nappe1]

No, each chip has a dedicated 20GB/sec bw bus. So, total is 80 GB/s of BW. The screen is divided into tiles, and each chip gets some of those tiles ("supertile" quote).

if someone didn't knew, afaik, Glaze as well as Axe did this same way.
Or other way to say this, quoting one of guys on BB 1.5 years ago: I can't wait to get my hands on the dual chip card... It will fly.

again boys were year ahead. Things like driving video input signal via PS unit and tile based multichip configuration on autumn 2001 would have been something I call nice.

okay okay... I'll stop now... my 9700 is still on package, because I am not yet back on my own comp.

 

[OpenGL guy]

I dunno, they don't list any bandwidth numbers, all they say is

Exceptional fill rates enabled by a 256-bit wide unified memory interface and integrated multi-chip supertiling

Unified.

There wouldn't be enough bandwidth to support 4 chips.

 

[BobbleHead]

I dunno, they don't list any bandwidth numbers, all they say is

Exceptional fill rates enabled by a 256-bit wide unified memory interface and integrated multi-chip supertiling

Unified.

There wouldn't be enough bandwidth to support 4 chips.

A single 256-bit bus for them all would also be unable to support the memory size required for "2-to-24-sample full-scene antialiasing at screen resolutions up to 2048x1536."

 

[megadrive0088]

with FOUR R300 VPUs working together, that means 1.3 BILLIONvertices/sec peak performance @ 300 MHz per VPU

325M verts/sec * 4 - I'm assuming that with 4 VPUs, it's not just increasing fillrate. it's fillrate, geometry and bandwidth. same with
last year's duel R200/8500 board for E&S.

with Voodoo2 SLI, you doubled the fillrate only, since Voodoo series lacked geometry processing (later called T&L, now vertex processing)

I assume it's similar to the way SGI's Infinite Reality series could be increased in terms of geometry and fillrate by using additional pipelines+rastermanagers boards

 

[megadrive0088]

I hope ATI designs the next Nintendo console chip, and that it is equal to
several R500s on one die, in terms of fillrate, geometry and bandwidth

 

[antlers]

with FOUR R300 VPUs working together, that means 1.3 BILLION
vertices/sec peak performance @ 300 MHz per VPU

From what was said above, geometry power does not scale. All the triangles go to all the chips, although the chips are able to quickly reject the triangles that don't fall into their tile.

 

[megadrive0088]

From what was said above, geometry power does not scale. All the triangles go to all the chips, although the chips are able to quickly reject the triangles that don't fall into their tile.

so it's for fillrate, not geometry and possibly not bandwidth either. bummer. well, it's not for consumers anyway, so we won't miss it. the simulation market demands incredible framerates anyway, which consumes fillrate. geometry isn't nearly as important AFAIK.


that does not mean though, that geometry and bandwidth could not scale with additional VPUs, as upto 256 can work together, correct?

 

[OpenGL guy]

so it's for fillrate, not geometry and possibly not bandwidth either. bummer.

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.

 

[megadrive0088]

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.

ok my mistake. so then that's two out of the three (fillrate, bandwidth, geometry)

80 GB/sec is pretty awesome. bet R400 won't have that. (40-60 GB/sec)

 

[jvd]

OpenGL guy ... what would the performance of that setup be in games ... or you can't say ?

 

[OpenGL guy]

OpenGL guy ... what would the performance of that setup be in games ... or you can't say ?

I have no performance numbers, but if you are fillrate/bandwidth limited, then I would expect performance to increase quite a lot I.e. close to 4x.

 

[jvd]

haha nice ... I wonder what the highest form of fsaa you would get with that card. I wonder if 16x would be playable at 1600x1200... that would be the card to buy....

 

[megadrive0088]

no worries, the next generation card from the team that made R300, the R500, will outdo 4 R300s in every way

 

[jvd]

Yea but thats what a year and a half away , if not more ? I want that performance right now . Its not a big deal if its only dx 9 . By the time i will need dx 10 or 9.1 or whatever they will have a card that is better and faster than mine and i will buy it

 

[g__day]

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. To get the expected ~ 80 GB/sec bandwidth wouldn't each GPU need the fully 128MB of memory, implying a 512MB card == very expensive. A card using only 64MB of memory per GPU would only have half that bandwidth (but still double that of a Radeon 9700 PRO).

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...