View Single Post
Old 12-22-11, 01:38 PM   #32
shadow001
Registered User
 
Join Date: Jul 2003
Posts: 1,526
Default Re: Official Radeon 7970 Review and discussion thread

Quote:

I'm surprised that no one has noticed these, even though they're theoretical but says a lot:








On both raw fillrate and especially texturing speed, it pretty much murders any single GPU card, but as always only when it's needed in an actual game and at high enough game settings.


Now notice where the biggest gains are for it generally?....That's right, in triple screen eyefinity where a lot more fillrate and texturing speed actually gets used up in practical terms, and why the performance gap is smaller when using a single display compared to a GTX580...The card is basically underused and severly limited by other factors( CPU or I/O perhaps?).


That's something that will be a constant even for kepler and still using a single display and expecting a huge performance increase....Using a single display isn't enough of a load period, no matter the game being benched.

It goes on further here:


Quote:


The most impressive thing is that the HD6990 has 64 Rops and the GTX590 has no less than 96 Rops in total, and they barely pull ahead in raw fill rate....So those have to be some mighty powerfull rops in the HD7970, since there's only 32 of them..


Same goes for texturing, both dual GPU cards have a lot more texturing units between both GPU's and the single GPU HD7970 has "only" 128 of them, yet again it isn't far behind when being mesured...


Gentlemen, those are some Rops and texture units with some serious real world performance right there, and a huge leap over previous single GPU cards, and it can't really be the 384 bit bus since the GTX580 has it too, and the HD7970 has about a 25% increase in memory bandwith over the GTX580, but some results are much higher than 25%, specifically when used in eyefinity...


This is one extremely efficient GPU period, then add the overclocking ability on top and lower power use and it's a pretty impressive package overall....




And more here



Quote:


Well that for single GPU comparisons once Kepler is out sure, assuming AMD doesn't increase the size a little further in a refresh part....Cayman is a 389mm^ part while tahiti is 365mm^, and even though 24mm^ doesn't sound like much, considering it's built at 28nm that's probably a couple hundred million transistors extra...


Where things really get screwed up is in memory bandwith, as the HD7970 is using the fastest GDDR 5 memory around and a 384 bit bus this time, so how does kepler beat that by enough margin in memory bandwith constrained scenarios so that it has a performance lead in those situations....Moving to a 512 bit bus perhaps?, and getting that extra 25% memory bandwith that way, since it'll likely be using the same GDDR 5 memory too.


Keep in mind that the true high end for AMD is the dual GPU HD7990 coming out in march, wich if the single GPU does what it does, crossfire scales even better with this generation, should give Nvidia engineers quite a few headaches in making a dual Kepler card to beat that one...We don't even know when the single GPU version of kepler will show up, never mind the dual GPU one to beat the HD7990...


Edit: i think AMD pulled an HD5870 with the HD7970 on Nvidia and it'll take them the better part of 2012 to release a full lineup to beat this, but that's just me.
shadow001 is offline   Reply With Quote