AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

[mc6809e]

Meh, until GPUs move to a new process, everything released will be completely uninteresting. It'll be better and faster sure - up to a point. But not very much faster or better. More than three years on the same process = extreme disappointment.

What blows me away is just how completely dependent performance is on process.

It's not that there aren't jumps with architectural improvements -- things like register renaming in CPUs and arrays of thread processing blocks in GPU, but it seems the muscle of process nearly always trumps cleverness.

It reminds me of the old 68K vs i86 battles. Without Intel's material engineers and process engineers, would they have won the battle? I doubt it.

Intel is now demoing processors at the 14nm node with up to 18 cores and 45MB of cache.

I wouldn't be surprised if they started trying to make equivalences between CUDA cores and CPU cores. "Why bother with GPUs when you can have a huge number of general purpose processors", they'll ask.

It's really too bad. A pure GPU in 14nm finfet would be a monster.

 

[silent_guy]

What blows me away is just how completely dependent performance is on process.

It's not that there aren't jumps with architectural improvements -- things like register renaming in CPUs and arrays of thread processing blocks in GPU, but it seems the muscle of process nearly always trumps cleverness.

Back when Kepler and GCN were introduced, with roughly the same performance/mm2 and, without reason, similar perf/W, I thought that both companies had come close to what could reasonably be squeezed out of a process. And then the GTX750Ti came, which doubled perf/W compared to its predecessors and improved performance by 30%. What blows me away is that they achieved performance and power efficiency increases on an already competitive device without doing a process change.

The next question is: how close are they now to the theoretical maximum?

 

[lanek]

for me its all about the rift. which likely means 90FPS @ ~1440/1600P.


so if this GPU is over 500mmsq what do we recon 64 CU?
if water cooled we could assume 1+ ghz. that would be a pretty mighty chip. I wonder how many CU's tonga currently has disabled?

We can assume reasonably that Tonga have just 1 CU disabled. and the full one is 2048SP.

Its a bit funny because if Tonga have been released at the same time of the Hawaii 290x, this card will have seems absolutely perfectly fitting this generation of GPU .. We can imagine the full Tonga will have got a bit higher/ equal performance of the 680 - 770 and just under the 290... basically showing for AMD a complete new lineup 290x-290-280x-280 , same as the 7970-7950-7870-7850... the high end with 512bits 4GB, the middle end with 256bits- 2GB ... I will have maybe release it directly with 4GB anyway...

 

[kalelovil]

We can assume reasonably that Tonga have just 1 CU disabled. and the full one is 2048SP.

That is what reviews of the R9 285 have been saying. However, if that is all that is missing from full Tonga, then it leaves unanswered the question of why Tonga's performance/mm2 is considerably lower than Hawaii and Bonaire (and Pitcairn).

 

[Alexko]

Is there any known difference in density between TSMC's 28nm HP and HPM?

 

[trinibwoy]

It's not just a density issue. It has a smaller memory bus and 700m more transistors. Question is what did those extra transistors bring.

 

[LordEC911]

Is there any known difference in density between TSMC's 28nm HP and HPM?

The better leakage/power characteristics of HPM would lead me, as a layman, to believe so.

 

[lanek]

The better leakage/power characteristics of HPM would lead me, as a layman, to believe so.

Better yields and so cost by waffer dont speak to anyone ?

 

[xpea]

It's not just a density issue. It has a smaller memory bus and 700m more transistors. Question is what did those extra transistors bring.

I only see new frame buffer compression and better geometry engines. seems poor for 700M tranies...

 

[Megadrive1988]

The generation after Pirate Islands is, apparently, called Faraway Islands:

http://wccftech.com/amds-future-gpu...ed-faraway-islands-features-20nm-coming-year/
http://www.tweaktown.com/news/39883...d-next-gen-faraway-islands-in-2015/index.html

So the first piece of information we know about this new family of cards is obviously the code name “Faraway Islands” or “FI” for short. The second piece of info is the manufacturing process and this one is interesting indeed. Instead of skipping 20nm and jumping straight ahead to 16nm which is what Nvidia is allegedly going to do, AMD is going to manufacture this family on 20nm and what’s more interesting is that it’s going to manufacture it at both TSMC and Globalfoundries. Cards based on this new architecture will also include support for Microsoft’s DX12 API

This obviously gives AMD the advantage of being first to market with a product based on a new process node which gives AMD a head start and forces Nvidia to play catch-up until it has 16nm products ready. Naturally the opposite is true as well and AMD would also have to play catch-up once Nvidia releases 16nm based cards. So we clearly see two different approaches here and only time will tell which will end up being the more successful.

The Faraway Islands GPU architecture will be integrated into all of AMD’s future GPU and APU products, including ones based on project Skybridge, so we’ll see this architecture in low power ARM based SOCs as well as good ol’ APUs which bodes very well for its potential power efficiency and scaling.

The fact that AMD intends to use a single GPU architecture in all of its products is also one of the possible reasons why it could not have skipped straight to 16nm like Nvidia is allegedly planning, since project Skybridge SoCs are all 20nm based.
What’s itching me most about this rumor and it is indeed a rumor is that CHW.com claims that

Faraway Islands an architecture coming after Pirate Islands is still 20nm based. What makes more sense for AMD is to either cancel Pirate Islands if it has a superior architecture on the same process node or scale the new architecture to FinFETs. And this is where my pinch of salt comes in handy.

 

[AnarchX]

I only see new frame buffer compression and better geometry engines. seems poor for 700M tranies...

+ 256KiB L2-Cache: http://www.computerbase.de/2014-09/amd-r9-radeon-285-frage-cache/
+ tuned VCE: http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/4
- ~7Gbps 384-Bit IMC

So this friday probably a 148W GTX 970 with GTX 780 performance will launch and in October could be a GTX 960 3GiB out, directly attacking Tonga.
So what did AMD in the 3 years after Tahiti design was finished? Try to keep up with NVs GPC-architecture from 2009 with a ~800 million transistor Quad-Front-End?

 

[Grall]

So what did AMD in the 3 years after Tahiti design was finished?

Designing, testing and releasing two console APUs may have consumed a lot of their internal resources.

 

[Kaotik]

The generation after Pirate Islands is, apparently, called Faraway Islands:

http://wccftech.com/amds-future-gpu...ed-faraway-islands-features-20nm-coming-year/
http://www.tweaktown.com/news/39883...d-next-gen-faraway-islands-in-2015/index.html

The problem with WCCFtech's theory is the fact that they can't get even the currently released stuff right.
Hawaii is Sea Islands just like Bonaire, Tonga is the first Volcanic Islands

 

[Psycho]

The generation after Pirate Islands is, apparently, called Faraway Islands:

No.. it was just a trap for certain sites... http://semiaccurate.com/2014/08/26/amds-faraway-islands-interesting-story/

 

[Alexko]

Come on, people, Faraway Islands is an area in Pokémon. Those people just read a joke tag from a SemiAccurate article and went wild with it.

Edit: too slow. But it wasn't a trap, just a joke. I don't think Charlie expected anyone to take it seriously.

 

[Megadrive1988]

Yeah, I just read that a little while after I posted.

 

[mczak]

It was obvious Faraway Islands was wrong, since Dave Baumann already confirmed the next series is named Fantasy Islands .

 

[Alexko]

It was obvious Faraway Islands was wrong, since Dave Baumann already confirmed the next series is named Fantasy Islands .

Hence the F.I. abbreviation! It all makes sense now!

 

[homerdog]

Scott at TechReport keeps going on about Tonga possibly having a 384bit memory interface. Is there any legitimacy to that theory? My gut says no, but my gut has been wrong before.

 

[Putas]

Well it can't be both 384bit and 1MB L2, right?