The ATI R600 Rumours & Speculation Centrum

So the door is open for 64 alus and 15-ish mb edram for r600!



...i kid

 

I believe I saw Dave B. state that their info was wrong. I could be wrong though.

 

Thanks S_P! I have been looking for that for some time.

 

Why would either of these upset ATI? Can you really "code" for a scalar architecture? I also doubt ATi will be too upset by a 32 pixel branching granularity - they've already been at 16 for goodness sake. Wouldnt even be surprised if R600 bests that.

 

Yeah, Wavey posted here fairly recently that these rumors of a disabled shader array were incorrect, and that Xenos incorporates other redundancy (presumably more fine-grained).

I'm slightly curious how that author is so sure of R600's shader count, even though consensus seems to be narrowing down to that, but that's a topic for another thread. Wait, that's the topic of this thread! Source?!

 

That video wont play!!! darn. Anyone have a link to a secondary?

 

ok found one that works for sure. I love you google.


http://video.google.com/videoplay?docid=-3583319920368555893&q=stream+computing

 

for you guys that want to listin to the 96 shader part, it on the 20 min mark. Just incase you dont want to listin to the whole 1 hour and 15 min presentation.


With that said,

Let me quote Dave O....

From the tone of his voice, it seems like he slipped out some R600 info.

 

The event is also semi-quoted (paraphrased?) here. To flesh out that purported slip (which has been decently discussed, in this thread among others):

 

Sorry, that's actually more generic than Tech Report's (apparent) mis-quote. They say "next generation" whereas he actually says "in the future" (yes, I just watched for myself as well). Well, in the future we may very well have 65,536 of them (because 2030 is as much the future as 2007).

http://www.techreport.com/etc/2006q4/stream-computing/index.x?pg=1

Edit: Ahhhh, a wee bit on he says "next generation" re "half a terraflop". So I think now that Tech Report combined those two thoughts. Of course, you also have to consider that the 65nm refresh is as much a part of "our next generation" as R600 itself.

Edit II: I was actually addressing Sound_Card --Pete snuck in between, the sneaky devil.

 

I like how he said at about 21-22 min mark, that their current generation (48 processors) can achieve 1/3 of a teraflop. Where the next generation (96 processors) can achieve 1/2 of a teraflop.

I guess you can derive some sort of performance ideas.

 

\


so it seems what Dave O. is hinting at is 2x across the board. He ties going from 48 to 96 from 3xx giga flops to 5xx giga flops. Which is R600's roumored target. He even adds 64gb/s to the words. Since it looks pretty much stone that R600 is going to be 512bit and will most likely have over double the bandwith because of it, it just looks like a possible lead up to what to expect.

 

Stick with "possible" and we'll shake on that and call it a night.

 

HA HA I always do.


But to chime in for a sec, dose not x1900xt have 48 pixel shaders and ?6? vertex shaders? That would make 54 shaders. 64..... even though unifide dose not look like a great leap. 96 xenos style ALU's sounds more right.

 

Actually, 48 + 8 VS. So 56 for R580. But of course it can't be that simple, because Rys and Uttar get way too much enjoyment out of laughing at us tech thickies for it to be that simple. For instance, R580 PS ALUs are 4D whereas Xenos are 5D. . . and that's before you even get to the idea that ATI probably tweaked those Xenos ALUs a bit. . .

I'm certainly not saying there isn't a case to be made for > 64 in R600. Of course there is, or I wouldn't be fence sitting.

 

You could be right about 64 while im right about 96 at the same time!


Their are two R600's. UFO and Pele. 64 sounds right for UFO while 96 sounds right for Pele.

 

Eight VSs, actually. Yeah, that's an argument for way more shaders. OTOH, we have to keep in mind that die space and transistor count are still limitations, and each "equivalent" (i.e., 4D:4D) R600 shader may be much larger thanks to I32 and all that general next-gen and specifically D3D10 jazz.

And I have to think space is at least a small factor given G80's "double-pumped" shaders. Very basically, they have 32 4D units operating as 64 when roughly "normalized" to the GPUs main clock. Given R600's apparent die size at 80nm vs. G80 at 90nm (the only "hard" leaked info available, AFAIK), do we think ATI has crammed in twice as many shaders at the same ("normalized") clock and offering the same general functionality? Will they compensate for the extra transistors with fewer TMUs? Would that be logical if they're really packing a 512bit external bus (and will it be 16x32b or 8x64b)? So many questions! I was leaning toward 96 shaders just b/c of Xenos, but Orton's comments haven't shed enough light for me to feel comfortable in predicting anything but ranges.

 

To double pump your shaders you have to spend more transistors making it faster as well. Deeper pipelines and/or more fancy fast logic. 10 pipeline stage ALUs for G80, how many for R5xx and Xenos?

 

Geo, I think I like your thought on 80 shaders!! It seems recently ATi do like a odd number on thier hardware

 

You could also try my (not patented LOL) layman's speculative reverse math. Assume you're thinking of =/>500GFLOPs, if you reverse that you can either have 64 dual-issue or 96 single-issue ALUs. I'd of course prefer the latter scenario, because if the 4:1 ALU:TMU ratio should be true, you obviously get more fillrate with the latter. Which is no absolute either, as long as I don't know what each unit is capable of.

Of course my puzzle would be way easier if one would tell me X GFLOPs and Y GTexels/s