R520 info thread #(9!)^9

Do ATI boards in CrossFire allow indepentant clockspeeds? I would assume so otherwise having the normal XT's faster would make no sense if the master cards were slower.

 

I suppose 600 instead of 625 on the core, and 1400 instead of 1500 memory means that CF XTs are "speed-binned" XTs.

Maybe it's to stop people buying the CFs instead of the XTs, so that ATI doesn't have to make too many CFs?

Jawed

 

So much for the 512MB version only being $50 more than the 256MB version.

£50 != $50. 512MB version costs $87.93 US more than the 256MB in the UK.

 

9AM EST, there they go...

 

X1800Xl and X1600 (P)review at hardocp

http://www.hardocp.com/article.html?art=ODIy

 

DH: http://www.driverheaven.net/reviews/r520reviewxvxv/

 

H and ET are up too. And B3D! Yay, Wavey! Rys is up at Hexus as well. Those boys are liable to be cranky today, so walk cautiously.

Top post updated here:

http://www.beyond3d.com/forum/showthread.php?p=586319#post586319

 

Damn DH is now crawling I had 5 more pages to load up, so i can see the whole review

 

320m transistors

Guess I heard that right then

 

What a waste of trannies for a memory controller!!!.If Xenos with the cache logic and more ALUs is near 250 million trannies!. And above all, by the benchs i have seen a 7800 with 550 mhz ( PS3 and G72 ? ) will smoke it.

 

No way! Next to the adaptive AA I find that the most exciting thing about this release!

It might not shine that much now, but it will when the R530 comes out.

 

I really doubt that all their caches before R5xx were direct mapped. Direct mapping in texture caches when you are accessing two different mipmaps doesn't seem like a great idea. In fact all the implementations published in the research literature have at least two banks (that's a 2-way set associative cache). Also going from direct mapped to fully associative caches looks like a very extreme move. In any case with most GPU caches being relatively small (unless they implement a larger L2 like NVidia seems to), not that high clock frequencies and no need for single cycle access fully associative caches shouldn't be that hard to implement in a GPU (compared with being impossible in a 3+ GHz CPU).

So batches (threads) are sized at 4x4 rather than at 2x2. That seems like something I will be testing next . 32 registers per 16 fragments means 2 registers per fragment which I found was degrading performance in the simulator even for UT2004 so I decided that 4 should be a better number. But their implementation and drivers may do a better job at fully using those registers (our current OpenGL ARB compiler doesn't even reuse free components in the constant bank, even less reuse components in the temporal registers).

 

Had this card made it out at the same time/before as 7800GTX, it would be a much better deal(seeing as there is no pounding). But then again I havent seen any overclocked reviews.

 

Blah. Just finished reading through [H]ard|OCP review. Too green for me.

Let's see another one.

 

Dear [H],

Since you're not using that X1800XT review sample, apparently, could you send it to Unwinder so he can make sure that R520 really only does have 4 quads?

Thanks,
Timmy

(I wonder if this will show up as a rumor on digit-life or whatever the constant pipeline rumor place was)

 

Any OC benches anyone ? Or aren't any tools there yet ? (HardOCP tested X1800XL against an OC BFG .. same price, they say.)

 

I'm also finding it a bit odd that I haven't seen any overclocking tests in any of the reviews I've read so far. I've already forgotten, was there OC tests for the GTX when that came out?

 

Yeah, which showed it clocking 70-100Mhz higher a lot of the time.

 

the more i see gibbo s posts the more i see it as we whine and bitch and exagigrate because we want to rape ure wallet

 

Take a look at the last page of our first part - bleedin' obvious how many quads it has.