AMD: R7xx Speculation

Status
Not open for further replies.
Wesker said:
Wait a minute...

If the clocks have been lowered AND the die size is around 256mm^2, then why does RV770 warrant an R600 styled cooler?

http://bbs.chiphell.com/attachments/month_0804/20080412_99a352fa5cdc00c12ef59TliiiYtmh6h.jpg

http://bbs.chiphell.com/attachments/month_0804/20080412_c081e23a80c02f4d6fdacVhZElGKEgGc.jpg

Either the die size is larger than what the current rumour mill suggest or that clocks are higher than what most sources are speculating.
Click to expand...


hmm the hd3870 needs a dual slot cooler too and it is smaller.
 
AnarchX said:
Such low? I thought to improve clock-speed was one goal of RV770 and the last number I heard was 875MHz for shipping.
Click to expand...

I've written about this before: it's delusional to expect much higher clock speeds in the day and age, unless you also accept a completely new architecture. Current assumptions are that RV770 is on 55nm just like RV670 and that it's an evolution of the same architecture. If that's the case, then there is simply no way the clock is going to be increased a lot. Even if it's the same architecture and on, say, 45nm, I still wouldn't expect much. All the numbers that are going around in the industry indicate that you'll save area and a bit of power, but the era of free speed upgrades is largely over. (Remember those utopian speculations about a 2GHz+ shader clock that some people thought was credible?)

One of the important points to realize is that it's not just 1 or 2 blocks in the whole chip that are limiting the overall clock speed. Closing timing at 800MHz with standard cells is *hard*, no matter which functional block you're talking about. Every designer of every block is designing against the same clock speed specification and all of them struggle to make it, iterating for weeks or months to shave picoseconds from the critical path. So upping the spec from 800 to 900 is a monumental affair that amounts to touching pretty much every block in the chip. For a company with limited resources (all of them) it'd be incredibly foolish to spend time on this, when they should be working on their next major architecture.
 
silent_guy said:
I've written about this before: it's delusional to expect much higher clock speeds in the day and age, unless you also accept a completely new architecture. Current assumptions are that RV770 is on 55nm just like RV670 and that it's an evolution of the same architecture. If that's the case, then there is simply no way the clock is going to be increased a lot. Even if it's the same architecture and on, say, 45nm, I still wouldn't expect much. All the numbers that are going around in the industry indicate that you'll save area and a bit of power, but the era of free speed upgrades is largely over. (Remember those utopian speculations about a 2GHz+ shader clock that some people thought was credible?)

One of the important points to realize is that it's not just 1 or 2 blocks in the whole chip that are limiting the overall clock speed. Closing timing at 800MHz with standard cells is *hard*, no matter which functional block you're talking about. Every designer of every block is designing against the same clock speed specification and all of them struggle to make it, iterating for weeks or months to shave picoseconds from the critical path. So upping the spec from 800 to 900 is a monumental affair that amounts to touching pretty much every block in the chip. For a company with limited resources (all of them) it'd be incredibly foolish to spend time on this, when they should be working on their next major architecture.
Click to expand...

Isn't that was Fast14 was for? :LOL:

I'd be very surprised if it clocked lower than the 3870. Hopefully they could just focus on improving the speed of the stream processors (and texture units if possible) and clock those higher.
 
silent_guy: Amen. The only thing I'm not sure I agree/understand your point with 45nm/40nm. I mean, that kinda goes against what I said in my latest 40G news piece - even if you optimized mostly for power efficiency, it'd still be very easy to get a 100MHz bump. Am I missing something here?
 
=>silent_guy: The rumours about RV770 having clock domains were pure BS, but the rumours about RV770 having a higher clock than RV670 had a basis - TSMC offers two variants of the 55nm process and RV770 should use the better one. I learned it from a source that doesn't know many details but wouldn't talk crap.
 
Lukfi, you keep repeating that, but I remain highly skeptical. There are two things to consider: public variants, which are highly generic, and 'real' variants that are much more precise. For example, there are apparently *12* variants at 65nm, which fit in 3 categories.

More precisely, those are transitor/gate variants, not processes; it is possible to mix-and-match some of them inside a single design if they were engineered for the same process characteristics (i.e. on 32nm, you obviously won't be able to exploit High Performance transistors unless you use high-k!) - I don't know the trade-offs or complexities involved in using multiple kinds of transistors for the same chip, but perhaps others would know better.

I can believe that RV770 would use a different transistor variant with both higher performance and higher leakage, but I am once again very skeptical. That'd require re-engineering the entire design for not that much fo a gain; plus, they presumably had a reason to choose that transistor variant in the first place...
 
Arun said:
I can believe that RV770 would use a different transistor variant (...) That'd require re-engineering the entire design for not that much fo a gain; plus, they presumably had a reason to choose that transistor variant in the first place...
Click to expand...
So, you think they would have to modify the whole architecture to "fit into" a better transistor variant? :unsure: Well, I have no idea how expensive that might be. But it would be logical for AMD/ATi to try not to spend too much on RV770, but get the best possible profit from it. That'd go hand in hand in trying to capture the most profitable market segment, so the focus won't be on performance, but rather on price/performance ratio. Maybe ATi doesn't want to compete with GT200?
 
I'd also imagine that ATI/AMD are rather gunshy of using yet another "leaky" process. And with the market they appear to currently be focused on higher performance at a risk of leakier transistors isn't something that would be in the cards.

I'm sure they'd rather continue the success of Rv670 rather than revisit the nightmare (power draw) that was R600.

Regards,
SB
 
Lukfi said:
=>silent_guy: The rumours about RV770 having clock domains were pure BS, but the rumours about RV770 having a higher clock than RV670 had a basis - TSMC offers two variants of the 55nm process and RV770 should use the better one. I learned it from a source that doesn't know many details but wouldn't talk crap.
Click to expand...

Eric Demers said:
Well, I think we have over 30 clock domains in our chip, so asynchronous or pseudo-synchronous interfaces are well understood by us
Click to expand...

If older chips had 30 clock domains don't you think newer ones would as well.
 
Yes, but the quote continues:
Eric Demers said:
In the R600, we decided to run at a high clock for most of the design
Click to expand...
Which would suggest that the R6xx architecture was not designed for running its parts (shaders) at clocks much higher than the rest of the chip. Although we don't know what modifications ATi did to the architecture with RV770, that's for sure.
 
Lukfi said:
So, you think they would have to modify the whole architecture to "fit into" a better transistor variant? :unsure: Well, I have no idea how expensive that might be. But it would be logical for AMD/ATi to try not to spend too much on RV770, but get the best possible profit from it. That'd go hand in hand in trying to capture the most profitable market segment, so the focus won't be on performance, but rather on price/performance ratio. Maybe ATi doesn't want to compete with GT200?
Click to expand...

I think it could be a good strategy. Speaking for myslf I want something that can run Crysis a lot better than my 9600GT. But I'm not going to pay 549 for it.

But reasonable prices in the 200-349 range as rumored for the AMD cards I might.

If Nvidia is going to do GT200 at a very high price and nothing else, at least initially (leaving only G92 for mid-low range, which should be trounced by the 4800 parts), as it looks, AMD could make a lot of hay. I think they might have a winner on their hands if 32 tmu's is true,
 
AMD To Launch RV770 On June 18th

AMD has just confirmed the marketing name for RV770 yesterday which is no surprise to anyone, the Radeon HD 4800 series. RV770PRO is Radeon HD 4850, RV770XT is Radeon HD 4870 and R700 is Radeon HD 4870 X2. AMD will be shipping RV770 GPUs to AIB by end of this month and you can expect different card designs from the manufacturers at launch. AMD is set to launch the Radeon HD 4850 on June 18th and it is definitely a hard one with retail availability on the same day. Radeon HD 4870, however, will be launched a week later on June 25th with retail availability in July due to GDDR5 availability. Radeon HD 4870 X2 will come later in Q3. As for the clocks, we won't be revealing them yet but the RV770XT core clock is not going anywhere above 800MHz.

Source: http://www.vr-zone.com/articles/AMD_To_Launch_RV770_On_June_18th/5777.html
 
Status
Not open for further replies.

Similar threads

F
RDNA5 Speculation
Replies
2
Views
1K
Erinyes
E
S
Apple Dynamic Caching on M3 GPU
Replies
5
Views
3K
Lurkmass
L
hoom
AMD: Zen 4, Speculation, Rumours and Discussion
8 9 10
Replies
199
Views
39K
MfA
M
xpea
Nvidia Hopper Speculation, Rumours and Discussion
7 8 9
Replies
177
Views
38K
Deleted member 2197
D
A
AMD: Zen 3 Announced [2020-10-08]
10 11 12
Replies
220
Views
91K
davis.anthony
D
Back
Top