R520 - talk, rumours, speculation; the lot!

Wishful thinking? I could care less, I assure you. But this is the quote in question:

I take this use of language to imply that the R520 as tested will be competetive with the GTX - but not exceed it. I could be wrong, but it's what I would interpret the statement as.

This is the exact quote from the admin:

I wasn't discussing it until you mentioned it, so no big deal - but I just want to point out that his reply of "8600 confirmed" was indeed to the 10vs24ps config.

http://www.hkepc.com/bbs/viewthread.php?tid=466843&extra=&page=3 (post #30)

I'm not harping on it, but nor am I going to just drop it as long as there's avenue for discussion there. For my part I'm happy just to wait for the R520/R580 and see. I've bought the vid card I intend to use for some time to come, so my interest here is purely architectural.

 

Exactly my point, its your interpretation (hence wishful thinking) that it will not be faster. HKEPC's statement didnt imply it would be faster or slower but you picked one.

Notice, he says confirm this "mark" only. (Nothing on the clocks & pp)

 

If this is the truth;

http://www.vr-zone.com/?i=2679&s=1

then, well how about: "Hey guys at ATI, YOU SUCK!"
They had almost a YEAR to tune drivers for R520 tech, and they still dont work well...

 

I chose not to interpret not just what was said - but what wasn't said. Why not say also that the 16-pipe version soundly beats the 7800? Why create a different sentence for it altogether saying it would be 'competetive?' Anyway - so that's my interpretation. Better to have one than not I feel.

Well, now *you're* the one thinking wishfully - because technically, the pipe numbers and clock rates were all a part of the sentence when he went to confirm 8600.

Certainly he could have removed that portion if he wanted. The author - from his article - clearly believes in the existence of 24 and 32 pipe versions; I don't know why you would rule out his making 3DMark claims for such configurations.

 

Performance seems about where I'd expect for an XL, maybe a little on the low side, but with room to improve. I don't know that I'd assume they have the latest drivers tho.

 

You are really stretching this ... :?

They have a card and dont even know the amount of pp. How could they suggest performance figures ? And their ending statements are their expectations.

I'll have to repeat myself .. notice he said "mark". Also he went on to point to it with an ASCII arrow.

All this is useless. You tried to take away from the article something you wanted, as clear as that.

 

Well, this is for the XL version. If you make a naive projection of performance in 3DMark05 you would end up with (~7000 3DMarks*(650/500)=) ~9,100 3DMarks for a 650 Mhz XT with the same pipeline configuration. Of course, 3DMark scores don't work like that, but you should probably also consider the 512 MB (v 256 MB) of video RAM, which I think will be a potent weapon (and more or less makes it obvious that Nvidia has to release a new 512 MB "standard" SKU) if that is the case.

 

Um, the 520 XL should be much cheaper than the 850 XT PE...it has more than 10% slower core and memory clocks....and at the same time is as fast or faster than the 850 XT PE.

Assuming the numbers are right....what's the problem exactly?

 

Oh, and what did I want to take away?

I think you're stuck in ATI/NVidia fanb*y world. Why would someone who wants the R520 to be 'worse' than the GTX also believe in the possibility of a clearly superior 24 and/or 32 pipe configuration part?

Your insinuations boggle the mind.

I extrapolated based on the exact reasons I stated - no more, no less.

 

Isnt R520XL supposed to go up against 7800 GT ? (7800 GT scores ~7200)

 

Ring theory

I was thinking about the ring vs. bus thing, and came up with a theory:

With a crossbar memory controller and 4 quad cores, each core has a 128 bit dedicated bus (512 total).
With a ring memory controller each core shares the 512 bit bus.

The big difference is that the result of a memory request is seen by all cores, not just the one that did the request. If all cores take the result of a texture request and place that in their caches it allows texture cache coherency between the quads.

See the previous solutions on texture caching:
1. R300 and up: no coherency in texture cache - different cores might make redundant memory access. Solution: use tiles large enough (16x16) so this problem is not significant.
2. NV40 and up: L1 cache has no coherence. Solution: L2 texture cache.

So the ring might actually make it possible for the R520 to distribute workload more fine-grained (quads instead of 16-tiles) without efficiency problems.

 

And why would they drop two quads at a time, rather than dropping from 32 to 28 or 24 to 20 (like G70)? That's the main reason why the rumors elsewhere seem less plausible than the rumors here.

 

http://www.vr-zone.com/?i=2666&s=1

I think it's not a repost.

 

That makes some sense.

 

Interesting thoughts, Hyp-X.

I'm expecting R520 to have a 16-way SIMD texture engine, like Xenos does.

I say that because if R520 shades fragments in a 16-way SIMD shader array (4 quads) then all 16 fragments will require texturing at the same time - hence the texture engine is also SIMD. That also implies that the texture cache is shared by all texture pipes.

So, while I've drawn the TMUs as four separate quad units (one per pair of memory controllers, effectively), I'm really expecting them all to dump the texture data they receive into a shared cache.

Jawed

 

Assuming both the rumors and my memory are correct:

Code:

R520: 16-1-1-1
|
 -- RV515: 4-1-1-1
R580: 16-1-3-1
|
 -- RV530: 4-1-3-2

I think Dave was replying to posts that compared [RV530:R520::RV350:R350]'s clocks and pipes, when RV530 seems to be a derivative of R580, not R520.

 

Man I would be really surprised if a 24PS 10VS part at 650Mhz only hits 8600 in 3dmark05.

 

Well, if true that would certainly provide some instant vindication against Serenity's incesent hassling.

@Trinibwoy: Remember that at that stage in the game, the XT would probably be CPU limited in much the same way as the GTX in the resolutions 3DMark tests at normally.

 

If is true, that R520XL 16p@500/1000 scores 7000, so 8600 could be result of 16p@650MHz R520XT (30% higher clocks -> 22% higher score - it can be possible: X850XT-PE is 35% faster on clocks than X800XL and 3D Mark 2005 score is 25% higher, it's the same case )