AMD: R7xx Speculation

Status
Not open for further replies.
Lux_ said:
I have gut feeling Jawed is right for most situations. But of course there probably are benches where RV770 reaches 2x. 3DMark06 for example ;).
Click to expand...

Well if it has 32 TMU's, there you go. Of course Jaws doesn't believe that. I believe he's just being stubborn. I dont recall all floated specs being majorly wrong on a card launch this close to the day before. And they all say 32 TMU. Are we also going to question the repeatedly floated GT200 specs? Nope, they're pretty well established. Same should go for Rv770.

Anyways, so the theoretical gets you close to a doubling, 100% more TMU, 50% more shader, add some clock increase..now it's not a straight doubling, shaders lag, so you might see less than 100. But again, 1.25 9800GTX averaged 82% 4870>3870 faster in the review I looked at. That's what I'm looking for..1.8-2X.

There's no way I see that with 16 TMU. If Jaws is right he has to be hoping for more like 20-30% improvement at most, which would be suicide. That doesn't even get you close to 9800GTX..
 
Rangers said:
Well if it has 32 TMU's, there you go. Of course Jaws doesn't believe that. I believe he's just being stubborn.
Click to expand...
I really and honestly hope you are right, but I'm really afraid the AMD is the stubborn one, playing it safe and keeping the changes to the minimum. This is regarding ALU clockdomain also.
 
Lux_ said:
I really and honestly hope you are right, but I'm really afraid the AMD is the stubborn one, playing it safe and keeping the changes to the minimum. This is regarding ALU clockdomain also.
Click to expand...

Adding TMUs isn't even considered much of a change. (Ding!)
 
Rangers said:
Well if it has 32 TMU's, there you go. Of course Jaws doesn't believe that. I believe he's just being stubborn. I dont recall all floated specs being majorly wrong on a card launch this close to the day before. And they all say 32 TMU. Are we also going to question the repeatedly floated GT200 specs? Nope, they're pretty well established. Same should go for Rv770.

Anyways, so the theoretical gets you close to a doubling, 100% more TMU, 50% more shader, add some clock increase..now it's not a straight doubling, shaders lag, so you might see less than 100. But again, 1.25 9800GTX averaged 82% 4870>3870 faster in the review I looked at. That's what I'm looking for..1.8-2X.

There's no way I see that with 16 TMU. If Jaws is right he has to be hoping for more like 20-30% improvement at most, which would be suicide. That doesn't even get you close to 9800GTX..
Click to expand...

That's what I thought also, it looks like most everything is doubled on RV770, which could mean it could be 1.25x9800GTX, which would confirm the rumors, AND which would make the pricing of $320 somewhat sensible.

Except z-fillrate, which is another bottlenecking point for the R600 and on ATI GPU's; and z-fillrate is only increased by clock increase?
 
Tchock said:
Adding TMUs isn't even considered much of a change. (Ding!)
Click to expand...

Meh, my last words what I believe rv770 might look like compared to rv670:
- 480SP (6x16x5)
- twice the z fillrate (per clock)
- still 4 samplers, but each capable of 8 TA + 8 TF instead of 8 TA + 4TF (btw - this high TA capability - was that also useful for AA resolve apart for vertex texture fetch / point sampling?)
- no idea about different ALU clock (I'm leaning towards a no). If true though I suspect sampler clock is the same as ALU clock.
 
annihilator said:
That's what I thought also, it looks like most everything is doubled on RV770, which could mean it could be 1.25x9800GTX, which would confirm the rumors, AND which would make the pricing of $320 somewhat sensible.

Except z-fillrate, which is another bottlenecking point for the R600 and on ATI GPU's; and z-fillrate is only increased by clock increase?
Click to expand...

Yes I forgot that's another HUGE, almost inarguable argument. As I say, companies will mess around with a lot of things but not price. Price is a rock hard thing. So just as it was proof HD2900 wasn't going to match 8800GTX the moment it's 399 price tag became clear, the 349 rumored tag on HD4870 also gives us an idea. It pretty much means it HAS to be faster than 9800GTX, and not just in a few "cherry picked" situations, or nobody will buy the thing, obviously. It makes the 1.3X 9800GTX argument fit in perfectly. And I dont see that possible at 16 TMU's (or if it is, I guess, AMD did well enough anyway).

The only caveat on this is if you believe the rumored 4870 MSRP isn't correct.

Personally I believe 32 TMU, 480 SP, no seperate physics, no seperate shader ALU domains. Clock looks to be around 850 but we dont have solid info on that.
 
If that's the case (850MHz clocks which aren't separate from ALU clocks) RV770 isn't going over the teraflops mark anytime soon, if its clock speed isn't over 1000. I don't know if AMD hyped about the teraflops thing, but I do believe the "Terascale" thing mentioned has to have something to do with that.
 
nutball said:
Well, without wanting to facetious or anything... if multiple CPUs was such a great solution why do we have 2GHz CPUs rather than 10 x 200MHz CPUs, and why were multi-core CPUs that last resort once clock scaling went wrong? If getting good performance out of multi-core CPUs was easy it would have been done already, and for CPUs multi-core is pretty much the same as multi-die-single-core when it comes to the software side (cheaper to build of course but that's integration for you).
Click to expand...

That is why I said if they can overcome some of the problems.

It is very analogous to CPUs like you said. It is far easier to do a quad core than a 8ghz core 2. If the software can catch up then it will in the end benefit people.

That was my point. If dual GPU configurations become more common then I believe some of their shortcomings are more likely to be overcome as there is incentive to do it. If they are only a tiny niche then why put the effort forth?

In no way am I suggesting this is good at the moment. I am saying in 5 years or more it may turn into a benefit.
 
Rangers said:
And they all say 32 TMU.
Click to expand...
What else do you want to fit into ~60mm² die size? 480 Shader-ALUs (remember, you'd probably need about 2,5 times the register-file too or you won't have much fun with your ALUs), 16 FP64-TMUs (including a fitting amount (~256 kiB) of L2 cache and improved point-to-point communication for X2-cards.

If AMDs engineers could do that, they'd make entry level chips with only 4 TMUs and 40 ALUs a LOT smaller.
 
Is there any way they are designing these chips with Spider in mind like to work better with amd cpus, they have these chips handle some of the processes handled by the cpu. so would a spider system be more efficient or are they designing it that way?
 
Quasar said:
What else do you want to fit into ~60mm² die size? 480 Shader-ALUs (remember, you'd probably need about 2,5 times the register-file too or you won't have much fun with your ALUs), 16 FP64-TMUs (including a fitting amount (~256 kiB) of L2 cache and improved point-to-point communication for X2-cards.

If AMDs engineers could do that, they'd make entry level chips with only 4 TMUs and 40 ALUs a LOT smaller.
Click to expand...
Don't think about diesizes, think about transistors, they have ~-220-240mil transistors to spend there if not more
Not everything scales in terms of diesize aswell as other things, so some things take relatively larger percentage of diesize on lower end products
 
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