AMD: R9xx Speculation

Well, it was worth a try…

 

5870 has ~20% more memory bandwidth than 4890, while having double the shaders, TMUs and ROPs.

They just can't increase memory bandwidth as fast as everything else.

 

It wasn't really necessary for Cypress.

 

The TMU performance is more influenced by texture cache bandwidth (and size) than memory bandwidth (*). The latter has a greater importance for ROP operations, especially blending, but that can also be offset by improvements (and better caches) there.

(*)
Idea: Double the bandwidth of the L2->L1 interface which stayed the same going from RV770 to Cypress (which halved the L2 bandwidth available to each L1).

 

Really disappointing. That should end the rumours about R6970 being faster than GTX 580. Again AMD will not show that so wanted faster single GPU card. ;puke:

 

There are 3 guesses about Cayman's 3DM performance on chiphell. You cherry picked the lowest one and present it as a proof of lacking performance...?!? Sorry, but after your irrational spamming related to Barts launch, you could (at least) think twice before posting such comments.

 

I'm:
a) Tired of the rolleyes smiley
b) Trying to see if you manage to contribute something to this forum at some point
c) Quickly figuring out that b) is a waste of everybody's time

So either you turn the ship around matey, or the current shall suddenly drive you to other waters. Ponder upon that.

 

But are these already scaled for fullrate 32bit INT?

On that point I remember comments by SirEric (I think) back in the days when ATI was doing FP24bit & he described how a bunch of the actual calculation steps were done at FP32bit but the other bits were only sized for FP24 because of die-size implications.

 

So, someone on chiphell forum takes Barts score of X1120 and for some reason adds exactly 50% (I'm guessing the PowerColor claim of 30-50% faster) to get X1680 and you come to conclusion that's an official score for Cayman in 3D Mark 11?

 

http://bbs.pczilla.net/thread-12510-1-1.html

 

Errm, why wouldn't they be? All of that is totally independent of data format, hence it doesn't matter if that's 32bit int or float, and the hw certainly can do full speed 32bit float...

 

So the new number is 28 simds? Noone mentioned 1792 SPs before, so I guess it must be true .

 

Or it's Cayman Pro...

 

That would be twice more SIMDs than in Barts.

 

28 SIMDs and 112 TMUs could very well be true. Basically Barts x2.

 

Well, Barts X2 would require 64 ROPS and 512bit bus, but I guess there would be a HUGE improvement over barts with the improved ROPS, even if the bus was still 256bit but with faster video ram.

It's a nice thought though that 6870 CFX is faster than the 580 and the 5970!

 

I hope you're a native chinese speaker/reader, as such chinese forums crossposts tend to introduce a lot of noise. For instance, are you sure what Google translates to "performance can not get 580" means "performance short of 580" and not "performance 580 is unable to reach"? And are you sure the source is not some random people with limited language accuracy himself? (like, say... worse than I, here)

28 SIMD would indeed make for a good HD6950 candidate, some ~20% faster than HD5870 depending on clocks (assuming 725~775MHz).

Unless AMD intends to promote itself with almost useless announcements multiplication and totally useless designs multiplication (or failed considerably), Cayman XT won't be there.

 

Or halve the count of L1s but make them dual-ported for TMU/ALU use?

 

And that reason is... simply... best case scenario... or something like up to X1680.

 

That new Chinese link says that "after respinning/reforging the card"... seems odd to suggest the card was just respin recently to get to 1792 SP's.

A little too close for release I think, esp. if they are indeed pushing the 6970 up to 12/7

So you really are pulling numbers out of no where and cherry picking posts to argue with people? Nice go at covering yourself...

What you posted here and what you posted up there are quite a different tone Actually, after your Barts posts, I shouldn't be surprised