What do you expect for R650

[trinibwoy]

Well Orton's comments regarding R650 sounded like it was just a shrink to manage power consumption but considering the context I guess it made sense that he focused solely on that aspect at the time. But depending on how scalable the architecture is and whether ATI had already encountered issues with R600 before starting work on R650 we could see some updates. Is the MSAA resolve issue a hardware bug or design decision? Also, I can't believe that ATI was surprised by R600 texturing performance - they designed it that way so it would be strange to see them change their ALU/TEX ratios. R600 is a formidable part already so I think if they can get R650 out with solid drivers and all the kinks worked out it will be very competitive.

However, they may have been surprised by G80 in several ways, not in the least by its gargantuan texturing ability. So that may have encouraged them to beef up those TMUs with an additional 8/16 filtering units each.

 

[Star_Hunter]

I don't know how ATI could be surprised that Nvidia upped their texture units from G70, and even if G80 kept the same number at 24 Nvidia would still be doing better texture wise. Why ATI thinks its a good idea not to increase it for two generations is just amazing. It seems like such a simple solution to the problem, but they might not address it until R700, and they better expect that G100 is going to see an increase as well.

 

[Per B]

ATI/AMD better concentrate on just getting the current R600 on 65 nm to lower power consumption and improve the yields. Maybe they will be able to up the clock as well. By doing bigger changes they will prolly be late again.

So I'm not expecting a true refresh, rather a R600@65. Should it still be named R650 then?

 

[Rys]

since it will be shrink at 65nm tech

Will it? If, and that's a fairly decent sized if, another big R6 chip shows up, don't automatically expect it to be 65nm.

 

[leoneazzurro]

Will it? If, and that's a fairly decent sized if, another big R6 chip shows up, don't automatically expect it to be 65nm.

Are you hinting that R600 architecture will be canned ASAP?

Anyway, this is what I would have expected from an "R650" based on a 65 nm process and plain R600 architecture:

- 96 shader units (5x) in 5 blocks (option 1) or 4 blocks with 24 shader each (option 2)
- Doubling of the filtering capacity, tweaks on the TMU, still divided in 4 blocks (option 1) or using more TU blocks (6 instead of 4) for option 2
- More ROPs (24)
- improved dispatch processor

 

[Rys]

Are you hinting that R600 architecture will be canned ASAP?

Absolutely not. The '6 family has some good legs left in it yet :smile:

 

[Mariner]

Absolutely not. The '6 family has some good legs left in it yet :smile:

When are we expecting benchmarks and more info about the RV610/RV630 then? :smile:

 

[silent_guy]

Or perhaps simply an R600 with twice the TMUs/ROPs count and a double pumped ALUs core?

As it is now, R600 isn't that bad and that would double raw power while not increasing transistor count too much.

You want to double the amount of everything that's not shader and MC. And you want to rearchitect the whole shader core more or less from scratch, because you can't 'just' double pump an existing design. But when you double the speed, you also double the relative latency, so all latency hiding related buffering needs to increase accordingly.

I don't see what's simple about it and how it would not increase transistor count much?

 

[nicolasb]

What I am hoping for in R650 is a lot of extra texturing power. I want something that can perform on a par with 8800GTX while rendering an outdoor scene in Oblivion with 16xAF switched on. (Not only is 8800GTX very fast at bilinear but it seems to take a remarkably small hit when going from bilinear to 16xAF - a mere 10% in one Oblivion benchmark I saw. Very impressive.)

What I'm expecting is just R600 again but with a faster clock and reduced power consumption. If there is actually a hardware bug in the MSAA processing, then maybe that will be fixed.

 

[aeryon]

Are you hinting that R600 architecture will be canned ASAP?

Anyway, this is what I would have expected from an "R650" based on a 65 nm process and plain R600 architecture:

- 96 shader units (5x) in 5 blocks (option 1) or 4 blocks with 24 shader each (option 2)
- Doubling of the filtering capacity, tweaks on the TMU, still divided in 4 blocks (option 1) or using more TU blocks (6 instead of 4) for option 2
- More ROPs (24)
- improved dispatch processor

and with so much changes, the part will come out spring 2008... definitively not a smart move with G92 already around the corner

 

[Megadrive1988]

I'm expecting the R650 to be only a speed bump/ very slight improvement over R600, not unlike R350 compared to R300. the biggest improvement will be from 65nm. smaller, cooler, cheaper. I don't think R650 will be an actual refresh-- that'll be in the form of an R680 or R700. I don't think we'll see more ALUs, textures, backends/ROPs until the refresh.

 

[Geeforcer]

Will it? If, and that's a fairly decent sized if, another big R6 chip shows up, don't automatically expect it to be 65nm.

So you are suggesting that if AMD optimise's the process to ramp the frequency to 1GHz+ while maintaining the similar power consumption at 80nm, that's what they'll do?

 

[dnavas]

I would think, from a strategic point of view, it would make most sense to play towards an apparent strength -- smaller and cheaper. While NV has yet to demonstrate 65nm, ATI has two on deck. If they can win the market where the majority of sales are made, they can choke NV off. Putting out an R600 at 65nm would theoretically be easier than pulling off an architectural shift plus die shrink. Presumably the next battle is for the back-to-school market. It remains to be seen what the R6x0 chips look like, but I think most will agree that the G84/6 were pretty 'meh' (at best). If ATI can refresh the R600 quickly at 65nm, I would espect them to do that, and suck the life right out of the GTS parts.

Not sure how much of a redesign it would take to add more bilerps to their TUs. I expect they would if they could, but winning the high end is icing at this point. ATI looks like it has a decent shot at doing well in the lucrative section of the market, and AMD is looking for positive cash flow....

 

[Geeforcer]

On the other hand, R600 turned out quite meh in many respects, and RV6xx chips are fractions of that original meh.

 

[nicolasb]

Will it? If, and that's a fairly decent sized if, another big R6 chip shows up, don't automatically expect it to be 65nm.

You mean it might be 55nm?

 

[pelly]

Do we know whether R600 is fabbed at one particular company or is it spread across multiple? (ie; TSMC, UMC, etc.)

One possiblity if it is fabbed exclusively someplace currently is the switch to another fab whihc has a better grasp on the process...(less leakage, etc..)

I'd be really surprise if there wasn't a die-shrink for R650 (if a refresh is made at all) since it seems the architecture is hungry for higher frequencies and lower heat...both which come typically with a move to a smaller process...

Time will tell...

 

[Kaotik]

Will it? If, and that's a fairly decent sized if, another big R6 chip shows up, don't automatically expect it to be 65nm.

I think they kinda confirmed that 65nm process at the Q&A session after.. hmm.. [something]690 lauch?

 

[leoneazzurro]

and with so much changes, the part will come out spring 2008... definitively not a smart move with G92 already around the corner

R650 is already in development.. and these are not really "changes", it's something like R580 vs R520.

 

[Tim Murray]

R650 is already in development.. and these are not really "changes", it's something like R580 vs R520.

Oh, really?

 

[leoneazzurro]

Oh, really?

That R650 should be already in development it's very likely (and hinted from many rumours), a GPU could not be designed and tested in 2 months, but I think you know that. If a R650 is not in development, then ATI days of leading graphics company are ended forever, plain and simple.

For the second point: I mean the architecture does not need to change drastically, not a leap like R420->R520.
R600 should be a fairly modular design, so adding resources should be not a big issue. It will be a different chip? Yes, of course. It takes no effort to do it? No, it will anyway require a good amount of work.

If yo are hinting something else, please clarify, I don't think I deserve sarcasm.