RV770 has 800 SPs and 40 TMUs from diagram????

[badgtx1969]

Saw this on VR-Zone, may be fake.
http://forums.vr-zone.com/showthread.php?t=289696

I still think RV770 will have 480 SPs and 32 TMUs, especially considering the sales web pages that were 'accidentally' posted today for the 4850. I would be awed and mystified if Ati/AMD pulled the ol' switcheroo, I guess we'll see on June 23rd. Looking forward to the RV770 launch far more than I was for today's GT200 launch.

 

[EXT64]

I don't know. The vast majority of claims I've seen said 800, but that link made me wonder some too. If it is a real 800, NVidia better watch out.

 

[iluvgillgill]

its 800/5 according to AMD how they calculate their shader, so its really 160 SP.

 

[badgtx1969]

its 800/5 according to AMD how they calculate their shader, so its really 160 SP.

Sort of, based on R600/RV670 each 'super shader processing unit' contains 5 sub units which ati calls shaders, these shader units aren't the same or as elegant as an G80/G92/GT200 stream processing unit which can perform more operation types and aren't specifically dedicated. You can roughly divide by 5 to compare the architectures, however, I believe ATi's shaders can do more calculations per cycle.

R600 architecture: http://www.beyond3d.com/content/reviews/16/8
G80 architecture: http://www.beyond3d.com/content/reviews/1/8

The question is, has ATi made any significant changes to their shader processing unit architecture. Probably only optimizations and tweaks (like G80 to GT200), but we can always hope for more

 

[TheGreatGrapeApe]

its 800/5 according to AMD how they calculate their shader, so its really 160 SP.

No, it would be 800 SPs.

How they are utilized depends on the dipatcher and arbiter, but it's no more 160 SPUs than the GTX280 is 10 or 30 SPUs due to their sub grouping.

Let me guess you're going to start calling the GTX SPUs 'cores' now too.

 

[TheGreatGrapeApe]

I still think RV770 will have 480 SPs and 32 TMUs, especially considering the sales web pages that were 'accidentally' posted today for the 4850.

That picture didn't show me anything that looked like 40 TMUs, and I think that number is a mistake of those reverse calculating a mis-print (swapped HD4870 fillrate with 4850 freq in early specs publishing).

http://www.tomshardware.com/forum/251632-33-rv770-800sp

I'm not even sure this confusion helps ATi, it would matter more if they were closer competitors, regardless of everything there's little else nV could do in response as they're already fast-tracking the 55nm replacement because of the 65nm parts own issues only a failure would change that, slowing down the rush maybe a little.

 

[JAYDEEJOHN]

Hmmm, after reading your link Ape http://www.tomshardware.com/forum/251632-33-rv770-800sp This leaves me still guessing. Is it 640 SP's? Did they change the ratio? Or even the composition/design of the SP?

 

[badgtx1969]

Ape, you are right it does look like 10 TMUs, however, a similar architecture diagram from R600 only shows 4 Texture Units. So does ATi cluster the TMU in 4's or is a TMU to them actually the texture filter units within each unit (the 4 pink/red boxes)?

Interesting to see that similar to R600, there are 80 ALUs per every 'Texture Unit'. Obviously there is no longer a ring bus w/ RV770 but there may be a crossbar/cache hierarchy that allows global data share?

R600:

Texture unit:

 

[badgtx1969]

wonder if this presentation shows an actual RV770 die shot, almost looks like 160 superscaler clusters

 

[radnor]

Nobody likes a troll, i know, But.

I approve this discussion. And will read more when i get to home.

 

[badgtx1969]

From ee times, AMD confirms that RV770 has greater than 500 'cores'


'AMD will claim technology leadership in two areas. Its chips will use more than 500 cores, more than double the 240 cores on the new Nvidia parts. They will also use GDDR5 memory interfaces running at about 3.2 Gbits/s or more. Nvidia will use the existing GDDR3 protocol running at up to 1.1 GHz on a 512-bit interface to deliver memory bandwidth up to about 102 Gbytes/s on some versions.'


http://www.eetimes.com/news/latest/showArticle.jhtml?articleID=208404063&pgno=1

 

[badgtx1969]

Hmmm, after reading your link Ape http://www.tomshardware.com/forum/251632-33-rv770-800sp This leaves me still guessing. Is it 640 SP's? Did they change the ratio? Or even the composition/design of the SP?

Maybe 4850 will be 640 SP and 32 TMUs from disabling 2 rows of the diagram? I am still on the hedge whether or not said diagram is real.

 

[DarthPiggie]

I guess the SP speculaiton on rv770 still haven't crash landed

 

[Amiga500]

Performance per die area.... 2.8x GTX200

The RV770 die is 250 mm^2.

How big is GTX260/280?

 

[Kari]

^^ 576mm^2, so rv770 ends up having over 20% more performance.... yeah..right.... X2 maybe..

 

[frostys]

And now after performance per watt, we now get performance per mm

Whats next???

 

[badgtx1969]

performance per TMU might look good especially if its 32

That presentation rhetoric is typical of AMD, never a raw #s comparison.

 

[JAYDEEJOHN]

Maybe this'll help http://my.ocworkbench.com/bbs/showthread.php?p=432681#post432681

 

[badgtx1969]

^Nice find, now how does the math work out for TMUs? Have to run, someone else can crunch the numbers.

 

[JAYDEEJOHN]

Its looking like 40

 

[TheGreatGrapeApe]

Can't reply, busy at work going to meeting, spent too much time in that 'optimize' thread.

Just a few things.

From my understanding of how it works, GPU-Z doesn't analyze the chip like Rivatuner, it simply reads the information being provided and displays it graphically. It's not detecting 800 cores, it's detecting a product code and reporting 800 cores, it reports the frequency and such from other inputs, but it's not actually detecting cores.

Edit out ringbus comment as was looking at he R600 pic when commenting (thought it was a reply to a reply), yeah I don't know about that one, seems strange they would abandon it, but since GDDR5 doesn't has a wire length worry you are a little freer in some respects but the crossbar to PCB interface would be packed and tiny on the 55nm chip. Doesn't look right.

Anywhoo, we'll see what's what, but definitely 100% that it's 32TMUs (even the early 3DD/Vantagemarks confirm it, but I don't know about the SPUs, I just feel that 480 or a similar 2^X number makes more sense than 800, which doesn't make sense without a major redesin in how it does things (like co-issuing or having something similar to the missing mul or more trascendental/branch/flow units) or more interesting would be making RBE components effective parts of the equation.

However IMO, people are confusing parts of this equation and reporting numbers and then people are assuming they mean the same thing.

I'll comment more later.

 

[The_Blood_Raven]

My head hurts...

 

[badgtx1969]

Can't reply, busy at work going to meeting, spent too much time in that 'optimize' thread.

Just a few things.

From my understanding of how it works, GPU-Z doesn't analyze the chip like Rivatuner, it simply reads the information being provided and displays it graphically. It's not detecting 800 cores, it's detecting a product code and reporting 800 cores, it reports the frequency and such from other inputs, but it's not actually detecting cores.

Edit out ringbus comment as was looking at he R600 pic when commenting (thought it was a reply to a reply), yeah I don't know about that one, seems strange they would abandon it, but since GDDR5 doesn't has a wire length worry you are a little freer in some respects but the crossbar to PCB interface would be packed and tiny on the 55nm chip. Doesn't look right.

Anywhoo, we'll see what's what, but definitely 100% that it's 32TMUs (even the early 3DD/Vantagemarks confirm it, but I don't know about the SPUs, I just feel that 480 or a similar 2^X number makes more sense than 800, which doesn't make sense without a major redesin in how it does things (like co-issuing or having something similar to the missing mul or more trascendental/branch/flow units) or more interesting would be making RBE components effective parts of the equation.

However IMO, people are confusing parts of this equation and reporting numbers and then people are assuming they mean the same thing.

I'll comment more later.

Thanks TGGA, I agree very confusing w/ so much differing speculation out there.

 

[JAYDEEJOHN]

Heres something "No, just 16 ROPs w/ 4Z/clk @ >100GB bandwidth" Does this explain it all, and howd that look and perform?

 

[iluvgillgill]

so when AMD claim to be more scalable and efficient where the RV700 got 800 SP compare to 240 in the GTX 280 thats more then 2.5 times. so should it perform 2.5 times better? from the review so far its not. so i guess its not more efficiently scalable compare to G200 then. am i right?