AMD: Sea Islands R1100 (8*** series) Speculation/ Rumour Thread

that's not true. both amd and nvidia did improve on 40nm for at least 3~5% in density. things like reduced vias, denser layouts etc could still amount to a few percentiles.
adding in the possibly reduced redundencies with 8xxx designs. we could be looking at up to 20% more raw horsepower with no gain on die size if all goes well
although 10% (not counting clock gains) would be more realistic. so 5b could still be under 400mm2, or 420mm2 could be more likely a ~5.2b chip

 

Which chips are you comparing, and how are those numbers derived?

An increase in transistor density from Cypress to Cayman based on transistor count divided by area could have been due to other factors besides an improvement in actual packing.
Devoting more die to logic means the share of area devoted to pads and I/O goes down. Since those parts of the die contribute more to area than transistor count, just bulking up for Cayman could have changed the density figures over Cypress without significantly changing the density of each type of circuit.

A noted change for Barts was choosing to forego memory speed for area savings, but that would be a choice that would not lead to an increase in performance as posited.

 

That's what everybody said about dual GF100/b. Now, we have GTX 590.

I would think that AMD is probably trying to "rush" Sea Islands in anticipation of the monster Kepler.

By the way, anybody here ever deduced that Barts was really VLIW4-based rather than VLIW5-based? I did a lot of empirical research on this, and it's the ultimate conclusion that the "knowledgeable" reviewers at sites like Anand and TR should have already figured for themselves.

 

Ya they only had to gimp the clock speeds by 20%.

 

And they never even melt or anything…

 

Nah, they just fried the eggs for breakfast.

 

Barts was known to be VLIW5 from the start. There were rumors to the contrary before launch, but it was a closed case already back then.

 

Source please? The only real source is a slideshow made by AMD, like the one that claimed Bulldozer to have 2B trannies (which AMD admitted some months later to have only 1.2B trannies)?

Realistically.. all right? I could start spouting all kinds of math to prove that it's VLIW4 if you only want to think that's it's a closed case without really thinking about it at all.

 

What sort of math are we talking about?


Highly anticipating Gipsel's response.

 

Eh, what? Chips cannot be rushed. They need their time to bake fully.

I'd like to know the details of your empirical research.

 

IIRC Gipsel had some documentation (prior to the 6xxx launch) due to which he was dropping hints here that Cayman = VLIW4 and Barts = VLIW5.

 

You might be getting new sig line material...

Edit- Uhoh

 

Y'all asked, so I started a thread on this!
http://forum.beyond3d.com/showthread.php?t=61594

Enjoy all of the number crunching... get your calculators out, ready and loaded!

I need the ability to edit my posts, though, to polish it, make it clearer to understand in a better format, etc..! Sometimes I need to refine myself!

EDIT: Thanks for this editing ability, to whoever it is! LordEC911, please don't mock me - it's a huge blow to my ego! J/k..

 

Teh more reliable source is directed tests, showing clearly an instruction throughput of transcendental functions scheduled alonside other instructions like MADD, that is just not possible with VLIW4, thus the 6870 is faster in this particular directed test than the VLIW4-based HD 6970.

And 5870 even bests HD 7970, because it can schedule a transcendental along four MADs if served right.

 

Thanks for being the only one here with a good attitude.

(to others)

Got a link to this? I'd appreciate it, thanks.

 

Sorry, no link except the one to the excel file on my usb flash drive. Not all truths are to be found in the internetz.

Short excerpt, so that there's at least something posted on the interent.
A roughly even mixture of a lengthy shader not doing anything useful with MUL, MADD, MIN, MAX and SQRT.

HD 5870: 1.206 GI/s. (Giga-Instructions per second)
HD 6870: 893 GI/s.
HD 6970: 877 GI/s.
HD 7970: 1.101 GI/s.


Might post the whole stuff later today on my website (no guarantees, though).

edit: To make it clear for further quotes (on some other forums... : The GI/s between Cypress XT, Barts XT and Cayman XT show clearly that the former two have additional ressources to execute scheduled transcendentals, i.e. the fat Rys- or T-unit, while Cayman and Tahiti have to execute those inbetween the other instructions, thus loosing throughput in this mix as was to be expected.

 

Actually, the hints were in the shader compiler (embedded to the driver). It was clear not the complete Northern Islands line could be VLIW4 (as there were only two ASICs with that architecture, the second one actually being the Trinity APU). But it is sometimes difficult to align with specific models (one can take the ordering of models/ASIC IDs of the past as a guide, but that changed at some point) without hardware at hand. So in the beginning (iirc the VLIW4 parts of the shader compiler appeared several months in advance of the launch in some driver versions) it was not entirely clear that Barts would be VLIW5.

To make a long story short, I got tipped off by someone else about Barts being VLIW5 and only Cayman being VLIW4. The day before the Barts launch I posted some code generated for the VLIW4 architecture and answered the question what this means, that one member of the Northern Islands family is VLIW4 ("It means that there is one member of a future GPU line from AMD which has indeed 4 slot VLIW units.", I later amended it means "at least one"). But a bit later someone in the thread spilled the beans about Barts being VLIW5.

 

http://semiaccurate.com/forums/showpost.php?p=155075&postcount=148

nice sharp .png image.. it would be more credible if it's taken through monitor by a 1MP camera

 

Now the picture of the future "2304 SP" card from Sapphire could have more sense...

 

Whats the Tflops on a 925mhz 7970? And what speed would the 7970's core need to be at to reach 4.5 Tflops?