ATI RV870, R870, R800

[thilanliyan]

http://www.nordichardware.com/news,8615.html

Cool if true.

EDIT (some more info):
http://bbs.chiphell.com/viewth...d=38229&extra=page%3D1

Don't know if it's fake though (the pics probably are).

 

[Denithor]

Story pulled.

:roll:

Should have quoted them instead of just linking...what was the gist of the article?

 

[Tempered81]

r800 is dual core (like an e8400) but will have two packages on one board, like r700.

end result is 4 rv870's on one pcb

BTW, I wrote this in a post in early December
2nd page:
http://forums.anandtech.com/me...id=31&threadid=2256125

Originally posted by: jaredpace
Well get used to it. The known next gen R800 is a dual core - dual chip PCI-E solution. Wether it be on one PCB or two, the high-end is still multi-chip and possibly multi-core MCM.
In the average of benchies, i can see this 260GX2 "GTX 295" beating the 4870X2.
The real surprise will hopefully either be NV's next gen on 40nm to compete with RV870, the future of raytracing/larabee, or lucid's hydra.

 

[Tempered81]

The story of a dual-die GPU from AMD/ATI has been circulating since way before the launch of R700, Radeon HD 4870X2. While it did allow many to dream, it turned out to be just a rumor, or maybe it was just to soon? Either way a source has hinted that the new card R870 will be two GPU dies glued together sharing a memory bus. This is pretty much the exact same that was said before the R700 launch, and while we can't confirm this, similar information has been posted over at Chiphell.

Both are basically saying that AMD has glued two RV870 GPUs together to build world's first dual-die GPU, which is kind of amusing considering AMD's bashing of Intel's dual-die solutions. Anyhow, this means that AMD would then be able to build what many would come to call quad-fire on a stick. Four GPU dies on the same card, R800. It also means that for the first time we will have a middle between the RVx70 and Rx00.

Specification wise this would land us on something like 1000-1200 shaders and 50-60 TMUs with RV870, 2000-2400 shaders and 100-120 TMUs with R870, and 4000-4800 shaders and 200-240 TMUs with R800. In theory (IN THEORY!), AMD could launch them as Radeon HD 5870, Radeon HD 5870 X2 and Radeon HD 5870 X4.

There isn't a salt mine big enough take with this, but we just wanted to notify you that the rumor has resurfaced. Beware.

 

[Bateluer]

So, with Quad-fire on a single PCB, will we gain the ability to run two such cards in a single system, for 8 effective cores. Octo-fire?

 

[thilanliyan]

Sorry...I didn't quote it.

Basically said RV870 is single, R870 is dual, and R800 is quad with possibly shared memory IIRC.

 

[videogames101]

Originally posted by: Bateluer
So, with Quad-fire on a single PCB, will we gain the ability to run two such cards in a single system, for 8 effective cores. Octo-fire?

And so it begins.

 

[Keysplayr]

So does this mean that all cores will actually share the same memory? If it's dual core,and there is 2GB on the card, it won't be limited to 1GB total buffer for each core? Or will it be the same thing it is now where each core will have half the memory on board?

Which will it be. Or other?

 

[nRollo]

Originally posted by: videogames101

Originally posted by: Bateluer
So, with Quad-fire on a single PCB, will we gain the ability to run two such cards in a single system, for 8 effective cores. Octo-fire?

And so it begins.

Agreed on the Octo fire- the diminishing returns we've seen going from 2>3>4 GPUs lead me to believe 4>5>6>7>8 would be even worse, not to mention the load balancing of such a beast.

I'd prefer a future with high end single GPUs and up to four GPU multi GPU configs.

I like the idea of four GPUs on a board but have my doubts about cooling for such a card.

 

[SlowSpyder]

Originally posted by: keysplayr2003
So does this mean that all cores will actually share the same memory? If it's dual core,and there is 2GB on the card, it won't be limited to 1GB total buffer for each core? Or will it be the same thing it is now where each core will have half the memory on board?

Which will it be. Or other?

I would think that multiple cores sharing the same pool of memory would be more efficient, assuming the bandwidth is there as opposed to 2-4 cores all using their own 1GB memory pool where everything is replicated.

 

[Denithor]

Originally posted by: nRollo
Agreed on the Octo fire- the diminishing returns we've seen going from 2>3>4 GPUs lead me to believe 4>5>6>7>8 would be even worse, not to mention the load balancing of such a beast.

I'd prefer a future with high end single GPUs and up to four GPU multi GPU configs.

I like the idea of four GPUs on a board but have my doubts about cooling for such a card.

I think it's going to be more like what's happening right now in desktop CPUs for gaming - build it and they will come. Seriously - look at a few of the newer games today and they actually support quad core CPUs for some impressive performance improvements versus duals (and single core systems - just say no). People grumble that GTA4 won't run well without a quad - to them I say that's where gaming is going, get used to it.

The same will happen with GPUs. If you make more cores standard, the devs will design their games with a higher degree of parallel multi-threading that will take advantage of the enormous crunching power available. It may take a few years but we're already seeing it - look at SLI/CF scaling today versus a year or two ago when the technology was newer & less common.

The biggest hinderance in my mind is that these multicore/multiprocessor cards aren't the standard solution (yet) unlike in the world of CPUs where practically every high end chip/system sold in the last 2-3 years has been multicore. The devs won't seriously scale their efforts toward this hardware until it becomes mainstream because mainstream is what drives sales. Honestly the biggest driver will be consoles - once these multicore or multiprocessor GPUs show up there games will be rapidly optimized for this kind of computational processing.

Originally posted by: SlowSpyder

Originally posted by: keysplayr2003
So does this mean that all cores will actually share the same memory? If it's dual core,and there is 2GB on the card, it won't be limited to 1GB total buffer for each core? Or will it be the same thing it is now where each core will have half the memory on board?

Which will it be. Or other?

I would think that multiple cores sharing the same pool of memory would be more efficient, assuming the bandwidth is there as opposed to 2-4 cores all using their own 1GB memory pool where everything is replicated.

GPUs need to follow the model already established by CPUs in this regard. IMC + small individual L2 + large shared GDDR5 (acting like the L3 cache on PhII+i7) will probably be the most efficient solution.