http://pc.watch.impress.co.jp/docs/2007/1204/kaigai404.htm

those who can read japs... pls translate...

lol

from the pix.. i'm assuming.. there is no more texture/etc/etc duplication in memory when running on XFire

This post has been edited by jinaun: Dec 4 2007, 11:05 PM

how come the japs got the news earlier? wheres the americans?

anyway r700 will be dual core huh?

This post has been edited by En.Vader: Dec 4 2007, 11:12 PM

its known r700 hav to be multicore oledi.. think of it as xfire on a stick..

its not feasible to make a single gpu with the target performance at current manufacturing technology.

But isn't GPU already a massively parallel processor? I mean, if you can put in multiple core on the same die (you must be able to cramp that much transistor on a die) why not come up with an architecture that has twice the SMP in one single core

I may be wrong but I finds multicore on GPU to be very strange..

ps: I'm not doubting the source. Just unsure of the benefits ATi sees in making multi core GPU.

Furthermore, what is the architecture? is it new, or are they just slapping in the same old architecture with smaller fabrication process and so cramping 2 cores on one die.

well.. i don't know whether they put multiple core into 1 gpu or a single board comprising of multiple smaller gpus

and.. to get more performance, you need to cram more transistor in the single die.. which currently we are hitting 700mill for 8800GT. i think its cheaper to manufacture smaller gpu compared to a single large gpu.

well.. its not entirely new, but i'm sure its based on R600 and new improvements incorporated to enable it to reach the target performance required for R700

maybe their sources are faster than the americans..

No, he's just speculating. The pictures are just speculation.

The transition of the GPU core is moving dramastically in next year. Looking forward to hearing from this in next year.

if it's true, then its good news indeed, because currently dual chip boards such as x1650xt gemini from sapphire have 2 sets of memory one for each gpu but basically both sets of memory holds the same data and this makes it redundant and it finally leads to an increase in production cost and pcb complexity thus increasing the price

This post has been edited by ben_panced: Dec 5 2007, 11:10 PM

View it as integrated cores...you know, like how Intel slapped 2 C2D to make a single C2Q. The integration have integrated memory controller that allows sharing of memory and is able to turn off a single core when working on desktop applications such as Office or the like. This is a replication of CPU architecture...but not sure if it's real or not.

Yes, GPU is already massively parallel. Just think of this as integrated Crossfire or something like that. No need for bridge, better buses, etc. Don't rely on this much though. I don't work with AMD.ATI, but I don't think this is where they are heading for R700...then again, I may be wrong.

It's not like a single dual core cpu sharing a single memory controller, because that would give them no benefit. It's more likely to have chips with independent memory controllers and nonoverlapping address spaces, like amd chips connected via hypertransport. It's an improved crossfire.

SIMPLE TRANSLATIONS:

First Title: Planned/ speculation of Future High-End GPU trend

Dual-die GPU: Mostly saying that the die will be half the size of now high-end GPUs, shared memory with lower latency, high bandwidth interconnection between MCM GPUs and thus lowering task processing overhead, blah blah...

Second Title: Power saving method on Dual-die GPU
Left Pic: High load
Right Pic: Low load

Yes, exactly what I'm trying to say. Seems like the merger is finally bearing fruits...but then again, it's a speculation as translated above.

This post has been edited by clayclws: Dec 6 2007, 11:48 AM

Let's just say, by the time the merger came about, R700 was well on its way already. AMD will have no influence on the designs till maybe the R800 or later generation. And i'm not saying it will use hypertransport, i'm just saying it's a similar configuration to amd chips using hypertransport. It probably will use a different connection method.

The diagram states its Multi Chip Module (MCM) so it's multiple dies (perhaps even > 2 ?) on a single substrate. Most likely will be connected via Hypertransport. It will be cheaper to manufacture the dies but slightly harder to assemble.

I believe Larabee is going to have multiple in-order x86 cores with separate L2 caches all attached to a larger shared memory: kind of like how C2D have a shared last level cache (LLC).

Wow, you seemed well informed. It seems logical though, since AMD themselves announced that their merger will only bear real fruit towards the end of 2008 and starting of 2009. Should be R800 or maybe RV7xx or something like that.

---

Added on December 6, 2007, 12:13 pmAnyway, let's just hope for the best for AMD.ATI. We need competition in all areas.

This post has been edited by clayclws: Dec 6 2007, 12:13 PM

CPUs are different, and i'm not expecting Larabee to be an out of this world GPU performer either. It will probably be "good enough" at the most for true GPU apps except maybe some triangle setup limited cases. They're probably going to market it as more of a gpgpu part and for "raytracing". And current GPUs are already multicore (have been for a while), with hundreds of threads per core. The problem is going multichip and getting good scaling.

Dual cores have existed awhile...i agree...the the dram sharing is it new?
And whats with the multi chip module? noob in this

Read what ikanayam wrote in this thread and you would know.