AMD: R7xx Speculation

Status
Not open for further replies.
Fudo doesn't know what he's talking about.
And, uhhh, 1000MHz memory? I don't think so, but we'll see...
Either way, yes, let's not waste our time in a thread about a very exciting chip to talk about a quite unexciting shrink ;)
 
Sunrise said:
Well, I don´t know how you guys feel, but right now I´m in absolute awe of the work that AMD/ATI have done behind the scenes. I just have to congratulate everyone at ATI that pursued an approach that was barely visible but (I see it now) definately is a great decision for the future.

I wish AMD/ATI success and many more customers, so that they can grab a big chunk of the market in no time and that they´ll continue to give customers great products for an unbeatable price (and hopefully still have their margins under control).

I must say that I really enjoyed the new tech demos/presentations, esp. the new ruby. That level of realism combined in one is just unreal (it´s only a 11sec clip for gods sake).

Now, speaking about the performance of RV770. I´ve never thought it´s possible that an ASIC of this size (I expected good, but not this) and on 55nm (just think about what ATI could do with the half-node 40nm) could deliver so much raw processing power. And we haven´t even seen what RV770 can do if you support it with enough BW, higher core clocks and it´s "new" multi-gpu-capabilities.

While RV670 really was a _great_ ASIC, RV770 is the king. Great work guys. Yes, Dave, feel free to forward that to your colleagues.

Looking at one of the previews that I personally trust the most, hardware.fr:

http://www.hardware.fr/articles/724-8/preview-ati-radeon-hd-4850.html

That´s only a HD4850 and it´s a summary of all games they´ve tested in 1920x1200 (with and without AA). If the HD4870 really is at least 20% faster all around, this would mean that we´d have about parity with the GTX280, give and take some %. This is "only" games of course. Can´t wait for the real McCoy (4870X2).
Click to expand...

It's about 40 TMU's..

And how the heck they managed to smash our expectations in BOTH TMU and SP count actually..

Basically, they seem to have a good deal more efficient architecture than Nvidia, suddenly. Packing a ton of functional units in little space somehow.

To me, it's not that a "monolithic" chip is bad, per se, it's that GT200's performance isn't good enough per die size. I dont think, if it was say 2X as fast as 4870 is going to be, people would be complaining so much about the 649 tag. GT200's problem isn't that it's monolithic, it's that it doesn't appear to be fast enough.
 
Rangers said:
To me, it's not that a "monolithic" chip is bad, per se, it's that GT200's performance isn't good enough per die size. I dont think, if it was say 2X as fast as 4870 is going to be, people would be complaining so much about the 649 tag. GT200's problem isn't that it's monolithic, it's that it doesn't appear to be fast enough.
Click to expand...
Bingo, agreed completely. If AMD decided to release a 600mm² chip based on R7xx and 512-bit GDDR5, the performance would be so mind-blowing it's not even funny.

All around though, the 65nm transition has been pretty damn shit for NVIDIA. There are rumours that G92b (and presumably GT200b too then?) will scale by more than the theoretical 19% though, so maybe that one will go a bit better. There's no way to get around the fact that NVIDIA lost their perf/mm² advantage going from 90 to 65nm though, while AMD created one out of nowhere with RV770.

What really matters now is whether G92b is 100% competitive with RV770 at good margins. If it is, then the financial impact on NV won't be anywhere near as big as some fanboys might like to think. At this rate, I'm getting skeptical about G92b though given RV770's stunning scaling in antialiasing modes...
 
Arun said:
Bingo, agreed completely. If AMD decided to release a 600mm² chip based on R7xx and 512-bit GDDR5, the performance would be so mind-blowing it's not even funny.
Click to expand...

That resumes all!. I completely agree with you. It´s not the today but the opportunities that RV770 bring graphic market for the future. I can´t avoid thinking in a 4xRV770 card at 1ghz in 40nm at 125w and more than 5 teraflops...
We got there at last, the way started by voodoo is near a fantastic ( and similar, hello voodoo 2,3,4...) end...
 
Last edited by a moderator:
randomhack said:
Tweak town reporting a new Nvidia card 9800 GTX+.
Looks like an OC GTX.
The prices : original GTX is now $199, GTX+ is 229 !!!!

http://www.tweaktown.com/news/9702/nvidia_surprise_with_9800_gtx_graphics_card/index.html
Click to expand...

From the French 4850 review, by game, 9800GTX versus 4850

ET Quake wars: Essentially tie, GTX faster without AA, even with
HL2 Ep 2: Essential tie
Stalker: tie
R6 Vegas: Decent ~20% win for 4850
Oblivion: Fairly close, 9800GTX bit faster no AA, 4850 significantly faster with AA.
GRiD: Solid 10-20% win for 4850
Bioshock: 20% win for 4850
CoH: Essential tie, GTX slightly faster no AA, tie with.
World in Conflict: Decent win for 4850, 26-19 FPS with AA
Crysis: Both unplayable, virtual tie no AA, 4850 19-12 FPS lead with AA

Now 9800GTX+ = ~9% overclock on shaders+core

I would say this move should stop Nvidia from just massively bleeding share, at least to 4850. 4850 may still have some other edges, such as more forward looking shader heavy architecture (theoretically) and DX 10.1. But it's not blowing away the new 9800GTX lineup, though it may be slightly superior, especially with a price edge on the GTX+. And the die sizes are similar? 4870 may still present a problem though. To which the umpleasent solution would seem dropping the 260 to prices much lower than Nvidia would like..
 
Last edited by a moderator:
Rangers said:
I would say this move should stop Nvidia from just massively bleeding share, at least to 4850. 4850 may still have some other edges, such as more forward looking shader heavy architecture (theoretically) and DX 10.1. But it's not blowing away the new 9800GTX lineup, though it may be slightly superior, especially with a price edge on the GTX+. And the die sizes are similar?
Click to expand...

I think it's interesting to watch if we're going to see the new GTX+ cards on retail or only in reviews.

Is NVidia already sending drivers which auto-overclock the GTX -> GTX+ to hardware sites so that when the 4850-reviews hit the net, it will be compared against the GTX+ -variant?
 
Arun said:
There's no way to get around the fact that NVIDIA lost their perf/mm² advantage going from 90 to 65nm though, while AMD created one out of nowhere with RV770.
Click to expand...
As for what happened to nVidia I was wondering about their current design flow. Apperently they put quite a bit of full-custom designed stuff inside G80 and later silicon, while this has given them an edge it has the drawback of requiring a redesign everytime you do something which is not a dumb optical shrink (and even an optical shrink is not exactly automatic). Automated flows on the other hand have improved significantly in the last few years, Fast-14 is just an example of that*, so I was wondering if AMD is reaping the benefits of improved tools while nVidia's approach has backfired.

This is obviously 100% speculation on my part but it could also explain the delays around the GT200 launch.

(*) Another proof of the significant improvements of automated flows in sub-90 nm processes is Cell's 45nm shrink. Even if it was far from optimal it was done by an extraordinarily small team in a very short time and with significant power and area savings.
 
Rangers said:
From the French 4850 review, by game, 9800GTX versus 4850

ET Quake wars: Essentially tie, GTX faster without AA, even with
HL2 Ep 2: Essential tie
Stalker: tie
R6 Vegas: Decent ~20% win for 4850
Oblivion: Fairly close, 9800GTX bit faster no AA, 4850 significantly faster with AA.
GRiD: Solid 10-20% win for 4850
Bioshock: 20% win for 4850
CoH: Essential tie, GTX slightly faster no AA, tie with.
World in Conflict: Decent win for 4850, 26-19 FPS with AA
Crysis: Both unplayable, virtual tie no AA, 4850 19-12 FPS lead with AA

Now 9800GTX+ = ~9% overclock on shaders+core

I would say this move should stop Nvidia from just massively bleeding share, at least to 4850. 4850 may still have some other edges, such as more forward looking shader heavy architecture (theoretically) and DX 10.1. But it's not blowing away the new 9800GTX lineup, though it may be slightly superior, especially with a price edge on the GTX+. And the die sizes are similar? 4870 may still present a problem though.
Click to expand...

Yes, but what about power consumptions ?
 
turtle said:
Ok, so breaking it down, I get a picture rounded out to look something like this (give or take):


1. GTX 280 = 4850x2 ($649 vs $400)
2. GTX 280 SLI = 4850x4 = 4870x3 ($1300 vs $800 vs $900)
3. GTX 280 3-way SLI = 4870x4.....both of which are bottlenecked... ($1950 vs $1200)

4. GTX 260 >= 4870 ($400 vs $300)
5. GTX 260 SLI = 4850x3 >= 4870x2 ($800 vs $600 vs $600)
6. GTX 260 3-way SLI = 4850x4 >= GTX 280 SLI = 4870x3 ($1200 vs $800 vs $1300 vs $900)
Click to expand...

I disagree with this assessment as I think you have the scaling off. A 280 in SLI is likely going to be qute a bit faster than 4x 4850's' and far more consistant, stable and reliable. Its the same with any of the comparisons using more than 2 GPU's. As for the 1 GPU vs 2 GPU comparisons. I think its fair to pay a premium to not have the headaches of dual GPU, i.e. lack of consistency, micro stuttering, input lag etc...

I'm not trying to big up NV or beat up on AMD, I'm made up that AMD are seemingly in a very strong competitive position again. i just think people are getting a little over enthusiastic and beating up on NV a little unnecessarily. Afterall the GTX200 series are still great GPU's, they are just a little over prices and that can easily change - especially in the case of the 260 which really isn't that bad anyway. $399 for a single GPU that slaughters an 8800Ultra?

I've numbered you comparisons above and taking into account the lack of >2 GPU scaling and the general problems with dual GPU's I would call them as follows:

1. 2x 4850 - this even makes me consider a dual GPU solution!
2. GTX 280 SLI - due to scaling issues it will likely be as fast or faster than the other setups in most situations while being far more consitent and stable.
3. GTX 280 x3 - It will probably be faster and again more reliable (although both solutions are poor choices IMO)
4. Until we see final performance and prices, I'm going to say this ones a tie.
5. As above between the dual 260's or 4870's. 3x 4850 is a loser IMO against those.
6. Probably the 280 SLI since it will be the most stable, consistent etc...
 
Miksu said:
I think it's interesting to watch if we're going to see the new GTX+ cards on retail or only in reviews.

Is NVidia already sending drivers which auto-overclock the GTX -> GTX+ to hardware sites so that when the 4850-reviews hit the net, it will be compared against the GTX+ -variant?
Click to expand...

I'm not sure about them sending auto-OC drivers around, but they are already sending out PR shit:

AMD’s Radeon HD 4850 has 800 shader cores. The GeForce 9800 GTX+ has only 128 cores. Doesn’t this put you at a disadvantage?

Core counts are as about as useful as MHz ratings. In the end, it comes down to real world performance. Our cores are designed to be exceptionally efficient and operate at near full utilization.

If you were to normalize 3DMark Vantage performance, per core, it would give you a better sense of the relative performance of different cores.
Click to expand...
 
CJ said:
I'm not sure about them sending auto-OC drivers around, but they are already sending out PR shit:
Click to expand...

wow.. last week having those 128 cores was like drinking wine alongside jesus!

Nvidia is about to start a new marketing gag. It doesn’t call Shaders, Shaders, anymore. As of today, Shaders will be known as Cores.


This is how Nvidia wants to communicate with its customers in the future. Intel has four cores in the desktop market and in 2009 it will likely go to eight cores; and Nvidia as of today has 240 Cores in its 1.4 billion transistor GT200 GPU.

What Nvidia fails to mention is that Nvidia Cores are not X86 capable and that they cannot actually run Windows and Photoshop; but Nvidia hopes that with adoption its CUDA marchitecture some application might adapt and work faster on Nvidia “Cores.”

This is a long, slow and painful process, especially when Intel as the main competitor has been pushing for X86 for more than thirty years. All compilers and software is written with X86 in mind.
Click to expand...
 
CJ said:
I'm not sure about them sending auto-OC drivers around, but they are already sending out PR shit:
Click to expand...

Lol, new efficiency index, performance/shader unit ratio
hihi.gif
 
Core counts are as about as useful as MHz ratings. In the end, it comes down to real world performance. Our cores are designed to be exceptionally efficient and operate at near full utilization.
Click to expand...
Pity the 64 TMUs and the 128 Z test units are working so inefficiently. Or the extra bandwidth.

Jawed
 
Status
Not open for further replies.

Similar threads

S
Apple Dynamic Caching on M3 GPU
Replies
4
Views
2K
Remij
R
hoom
AMD: Zen 4, Speculation, Rumours and Discussion
8 9 10
Replies
199
Views
35K
MfA
M
xpea
NVidia Hopper Speculation, Rumours and Discussion
7 8 9
Replies
168
Views
32K
pharma
P
A
AMD: Zen 3 Announced [2020-10-08]
10 11 12
Replies
220
Views
87K
davis.anthony
D
Rys
Reopening Architecture and Products + rules and guidelines changes
3 4 5
Replies
90
Views
13K
Shifty Geezer
Shifty Geezer
Back
Top