RV775XT and PRO have 840 SP?

[w0mbat]

Maybe, maybe not. There was talk about more than 800SPs many times before this.

[NBF]

Interesting theory, not the first time I hear it. But this time there is a picture

Goes to the news

[cadaveca]

Are you serious?

[Calmatory]

Are you serious?

Well, everyone could just as well shutup and wait till the thing gets launched, if you prefer that.

[Voodoo²]

From Xbitlabs:


This is from june.

[ColonelCain]

Here's my guess for "RV775" (= cherry picked RV770).
10 SIMDs
900 ALUs
40 TMUs

Yep.
90 ALUs per SIMD.
Cause that's what I see when I look at the die of RV770. There's nomore than 10 SIMDs but each of the ten SIMDs has infact 9, and not 8 clusters. No redundant TMU exists on RV770.

I was wondering how long it would take someone to look at the die shot...

It's also possible a respin removed UVD hardware and replaced with ALU clusters, but not likely.

"RV775XT" is the AMD equivalent of GTX260 216......TSMC's process has matured enough for such as show by nV already...

Hey! I was the first one to see that! No fair! (Don't mind)

10 rows and 9 identical blocks in each. The "9th block" being the redundant part would be my guess.

Attachment 91041

Thanks for the picture largon. I actually spent a good chunk of time researching this rumor to see if it had any substance. And, when I looked at the die shot, I noticed the extra group of shaders. Was thinking of doing a quick paint edit and posting it, but had to resume studying for finals.

From Xbitlabs:


This is from june.

This is the picture that I saw.

Anyhow, for you guys whom can't see it:

Look closely. If you think of it like this: Each "corner" structure is part of an SIMD block, you get 4. X5 (for the 5 shaders in each unit), then X4 for the 4 blocks composing each SIMD core. Then X10 for all 10 SIMD cores. You get 800.
And, for those that need an even bigger view, here's an 861 x 848 image that I found (but chose to leave it as a url, so people wouldn't complain about it being too big).
http://upload.wikimedia.org/wikipedi...0_Die_Shot.png


Anyhow, if you compare the 1st and 9th rows they look nearly identical. Perhaps, they are different because there needed to be a slight design alteration to put in a half "block"? Do I know that that is the exact reason? No. But one thing that I am fairly sure of is that the "structure", is the same as the shaders to its right.

Lastly, for anyone who don't know where largon got the 900 number from, it looks like each of the 10 "half-blocks" in the 9th row have 2 units each. 2x5x10=100.

I don't want people to get the idea that I am trying to give a math lecture on simple adding and multiplication. I just am laying it out there for anyone that is confused.

[cadaveca]

BTW, that block diagram from XBIT isn't exactly correct, but whateva.

[Helmore]

That block that you refer to ColonelCain is not as similar as the other blocks as you think.
Here is another bigger die shot:
http://picasaweb.google.nl/lh/photo/...cTvr9zQQrg1_Xg
Hope the link works like that for everyone. You can download the picture from this link if you want, that may make zooming in easier although the conversion from png to jpg has decreased it's picture quality .
This extra block might have something to do with scheduling or something, but I don't think it's some sort of extra shader unit.

[Monkeywoman]

Xtremesystems has done it again!! look over here; http://forums.vr-zone.com/showthread.php?t=368145 !!!!!

and then look over here http://www.xtremesystems.org/forums/...&postcount=120 !!!!!!

[tenebre]

That block that you refer to ColonelCain is not as similar as the other blocks as you think.
Here is another bigger die shot:
http://picasaweb.google.nl/lh/photo/...cTvr9zQQrg1_Xg
Hope the link works like that for everyone. You can download the picture from this link if you want, that may make zooming in easier although the conversion from png to jpg has decreased it's picture quality .
This extra block might have something to do with scheduling or something, but I don't think it's some sort of extra shader unit.

the 9'th column is different

[STaRGaZeR]

Man largon what have you done...

[largon]

Whoa...
And all I said was "Here's my guess..." and "...would be my guess".

[Tiridum]

zoom into that picture, the 9th row of "shaders" looks different from a regular row, the light pink things are higher and there is a dark pink column missing between the light pinks things and the texture units

[enteon]

I think they're different, too.

and it wouldn't make any sense for AMD not to activate all SIMDs on their very expensive and presumably high-quality FirePro V8700 which, in fact, only has 800. If even the most qualitative chips don't come with more than 800 usable SMIDs then there simply aren't more.

[STEvil]

maybe those are the parts for the currently disabled sideport?

edit - or power controllers to enable/disable shader arrays for power savings.. just didnt work out on 4870 maybe?. woo theories.

[cadaveca]

ONly half of the claimed parts are SP cores...the bottom half. It's not 900SP, hence me laughing about this being news earlier.

[RejZoR]

It wasn't really a news and it wasn't related to RV775 series. I found it somewhere else and just asked here for confirmation if that is true. It turned out it is, partially.
But nothing end users can actually use.

[renzokuken]

At the risk of sounding noobish (but in the spirit of learning!)...how important are SP's? I never understood how, for example, the GTX 280 has 240 SP's while the 4870 has 800 (!) and yet the 4870 isn't leaps and bounds better than the GTX 280 (in fact it's worse despite having more SP's and a higher GFLOPS output). Where does the "problem" lie in ATI's cards (the architecture?) and why can't we compare the performance between the two cards based on the # of SP's in ATI cards to the # of SP's in Nvidia cards? If anyone can answer either of these questions it would be greatly appreciated.

[hurleybird]

I never understood how, for example, the GTX 280 has 240 SP's while the 4870 has 800 (!) and yet the 4870 isn't leaps and bounds better than the GTX 280 (in fact it's worse despite having more SP's and a higher GFLOPS output)

GTX 280's SPs are clocked much higher, and, on average, operate closer to their theoretical power than the SPs in R700.

[Monkeywoman]

ATI SPs are locks to the same speed as the core (4870-750/750/900) with Nvidia they are separate (8800gtx-630/1000/1350)

one ATI SP can do 5 things per clock where as the nvidia can do less but it equals out to be more because of the shader clock AFIK

[renzokuken]

GTX 280's SPs are clocked much higher, and, on average, operate closer to their theoretical power than the SPs in R700.

Awesome! Thanks so much for the exact answer I was looking for.

[Macadamia]

At the risk of sounding noobish (but in the spirit of learning!)...how important are SP's? I never understood how, for example, the GTX 280 has 240 SP's while the 4870 has 800 (!) and yet the 4870 isn't leaps and bounds better than the GTX 280 (in fact it's worse despite having more SP's and a higher GFLOPS output). Where does the "problem" lie in ATI's cards (the architecture?) and why can't we compare the performance between the two cards based on the # of SP's in ATI cards to the # of SP's in Nvidia cards? If anyone can answer either of these questions it would be greatly appreciated.

Let's put this way.

ATI's architecture has 160 Execution blocks. Each of the blocks can do 5 ops simultaneously if under an ideal condition in VLIW. In most cases 3-4.5 per 5 "shaders" is used.


nVidia's GT200 architecture has 240 ALUs that can do dual issue. They also run at 1+GHz clock speeds, which makes them faster in processing.

However, there are a few notes to consider:
1. GT200 has 80 TUs. RV770 has 40 (32 interpolators 40 filters).
2. GT200 has 32 ROPs. RV770 has 16 (which are on par until 8x MSAA, where they take the lead in Z-fill)
3. GT200 is almost 2x the size of RV770 with both in 55nm (480 mm^2 vs 256 mm^2)

It's not as simple as you think. Texturing power still plays quite a role in graphics performance, but I do predict that once you have ~64 TMUs (750Mhz quantified) you've pretty much reached the peak.


I'm pretty confident in ATI's next chip on 40nm. It's not a performance evolution per se, but from a silicon budget perspective it should be quite crazy, making even the RV770 look bloated (let's not even mention GT200 here).

[ColonelCain]

ONly half of the claimed parts are SP cores...the bottom half. It's not 900SP, hence me laughing about this being news earlier.

You sound like you are very well versed in the topic of GPU uarch. Could you please elaborate? I'm very curious.

[cadaveca]

LoL. ME? Well versed? I may sound like I know what I'm talking about, but please take ALL MY POSTS with a grain of salt...


Really....Thanks, but I don't know as much as my posts seem to relate. I mean, sure, I've been following gpus for years, and it's the silicon that interests me really, and what made me an ATI fanboy, as ATI has had "new" tech in every gpu, since R300 was designed for M$ to test DX9 on...but I only know what is related by the company itself, and I listen to investor meetings and such stuff, as sometimes little tidbits fall through....but I know NOTHING 100%....probably not even 50%...

Anyway, given that I know that the sideport is a pci-e port(next to the other pci-e port, STEvil), and given that although SPs are right there, so are other bits that need some redundancy as well, and understanding that a bit of non-active silicon right in the middle of the gpu would help deal with heat issues, only so many things are possible.

Plus add in the XT naming that came with this news, and knowing who used that nomenclature in the past, I think I know who sourced the info, and I know they'd be wrong.