AMD R600 to use GDDR4 only

[Ziwro]

IF YOU REMEMBER the GDDR-4 story back in September, Samsung was telling everybody that GDDR-4 is going to rock. However, companies were worried that yields are not going to be good, and rumours about Samsung screwing up forced affected companies to spend thousands of dollars to invest in the development of alternatives.

In the end, Nvidia went with GDDR-3 memory in its G80 line-up, but DAAMIT continued to develop both GDDR-4 (Pele) and GDDR-3 (UFO) variants. Being late on the market enabled ATI to optimise. However, bean counters at AMD recently gave a "go ahead" signal for full-GDDR-4 line-up. So, forget about the dual-SKUs, 1Gig and 2Gig versions, GDDR-4 is the only way DAAMIT will go .

Also, the product formerly known as FireSTREAM, conveniently renamed the AMD Stream Processor (GPGPU board) is also in the works. We are unsure of the name for this R600 based one, but one of our sources was caught saying: "With their level of imagination, I don't see why not to go AMD Stream Processor 2 - The New Degeneration". If AMD will be gunning for the price, introduction of 2GB UFO board sounds perfectly reasonable, but our sources are telling us that the GPGPU board is a GDDR-4 baby as well.

Memory layout remains the same, as does the GPU. The chip itself is rotated at 45 degree angle – not the 60 degree this tired hack wrote a week ago. Also, the chip whose pictures were leaked was not a functional chip, but rather a mechanical "dummy".

The traditional layout of ATI boards calls for DVI-DIN-DVI arrangement. However, this might change and R600 could end up with G80-style DIN-DVI-DVI, with the HD connector located at the top of the PCB, not in the middle. Just a bit over two months to go

Orginal Source (http://www.theinq.com/default.aspx?article=36091)

[Revv23]

the R600 really gets my blood going.

[FghtinIrshNvrDi]

I'm ready to be teased by them with some leaked benches.

Ryan

[turtle]

I'm ready to be teased by them with some leaked benches.

Ryan

*sigh* Yeah, me too. Shouldn't be long now. :)

[metro.cl]

GDDR4 fr that time wont be that expensive.

Nvidia will have a vga with more shaders (160 is my bet) and GDDR4 to counter r600, i got the info that nvidia can add easily more shaders at nvidias launch for southamerica a couple of days ago

[turtle]

10x16 eh? Seems like the next logical step.

Perhaps nvidia can add shower power so easily because it's hidden in the inactive MUL?

I kid.

Why not 12x16 (192) and give each array it's own 32-bit bus? That'd be sweet.

[metro.cl]

10x16 eh? Seems like the next logical step.

Perhaps it's hidden in the inactive MUL?

I kid.

Why not 12x16 (192) and give each array it's own 32-bit bus? That'd be sweet.

yes that would be sweet but maybe a bit high, i still dont buy all r600 rumored specs

[turtle]

Which one? I don't trust them either, as there are so many different ones.

Oh...you probably mean the 16x32 512-bit (16x[32x2] internal) bus....yeah, i'm not quite sold on that one myself...:p:

[FghtinIrshNvrDi]

Other than the ~10800 in 3DM06 on a FX-62 that's floating around?...

That's news to me. With my 4400+ clocked at 2.6, I sorta have an FX-60, so I spect that would land me ~10,000ish... assuming that's the real deal.

I'm ready to see some price battles. Geez, I need to save up some cash.

Ryan

[dinos22]

Other than the ~10800 in 3DM06 on a FX-62 that's floating around?...

*sigh* Yeah, me too. Shouldn't be long now. :)
do share links with us all :toast:

[DilTech]

do share links with us all :toast:

I've heard the same thing at beyond3d.

[GrimReaper]

Other than the ~10800 in 3DM06 on a FX-62 that's floating around?...

*sigh* Yeah, me too. Shouldn't be long now. :)


yeah post the links of this

[biohead]

*slobbers*

[kiwi]

Without links or any believable screensot or pics I can say r600 scores 14000 in 06 :)

[w0mbat]

Yes, I heard the R600 ES (with lower clocks) @ FX62 @ 3D06 = 10800 points too.

[Ziwro]

Without links or any believable screensot or pics I can say r600 scores 14000 in 06 :)

Thats good :D

[RPGWiZaRD]

GDDR4 was quite expected since it was delayed this much and all. Nvidia 8900 series will most likely also all be GDDR4.

[madcho]

Yes, I heard the R600 ES (with lower clocks) @ FX62 @ 3D06 = 10800 points too.

thx for the news :toast:

[nn_step]

R600 is gonna be sweet :slobber:

[Cuthalu]

Yes, I heard the R600 ES (with lower clocks) @ FX62 @ 3D06 = 10800 points too.

We're all probably heard the excactly same thing, and none of those are "confirmed" except 10800, which as well might no be true. :stick:

[nn_step]

We're all probably heard the excactly same thing, and none of those are "confirmed" except 10800, which as well might no be true. :stick:
what makes you think that?
More time to design usually means better scores

[Cuthalu]

what makes you think that?
More time to design usually means better scores

I didn't comment on that. I commented about what actually is known and what is only guesses. According to that source.

[Shadowmage]

Link to the 10800?

[Pinnacle]

Link to the 10800?

Dont believe it, ppl are just pulling numbers out their asses

[turtle]

Dont believe it, ppl are just pulling numbers out their asses

I'm not denying that it could be crap, not at all.

[turtle]

I want to hear more about the MSAA capabilities. Fast video cards are nice, and winning or losing benchmarks is lovely, but image quality (with the least performance detriment possible) is what excites me...Hopefully they raise the bar again.

[Cuthalu]

Umm, Conroe's MHz is faster than FX's.

[generics_user]

at same clocks the FX62 would surpasse teh conroe,
at 2.93ghz the conroe scores only 500points better.
just checked the orb and difference between stock fx62 and stock dudo are about 1400 points:fact:

[kemist]

The 512bit bus rumor looks more likely; they got a picture of the g80 pinout and there are only ~1450pins compared to the about ~2000 of the R600.

[turtle]

1449 to 2140. :D

But who knows...I thought the dude whom mentioned that the outter pins being SBGA was onto something, as that would likely be the ring bus. I don't know how it compares to the fraction used in R580's 1265 pins , but it would seem to make sense.

http://img88.imageshack.us/img88/935/goohl0.jpg

My 5-second paint job doesn't outline exactly the right pins (one shorter indented row above and below my box), but i'm sure you can tell the difference.

[largon]

turtle,
Afaik, ring bus is an internal datapath, it has no external contact points.
The idea behind the ring bus is that the memory controller is closely coupled with the ring bus offering fast DRAM access to all parts of the GPU.

[turtle]

I must have interpreted this wrong:

http://www.extremetech.com/article2/0,1697,1867243,00.asp

I was under the impression the IMC was fashioned in a circular method around the outside of the die, with 16 32-bit channels, w/ traces going to the (presumably 8 x [32-bit x 2]) ring stops. I thought this perhaps a reason why the die turned 45 degrees, to allow an easier way to equalize the traces between the greater amount of ram (16) and the larger amount of of "bus stops" where the channels (MC to RAM) are connected (8). My understanding was that in R5xx's architecture, it was approx half that (8 32-bit channels, 4 32-bit x 2 ring stops). If that is incorrect, could someone clear up how it does work?

Any ideas why those pins are different? It may not be for the memory bus, but obviously a external 512-bit MC does need a fair amount of pins, so I tried to put it together.

[Cuthalu]

Where did Razor say that?

[razor101]

Hey this is Razor1 from B3D, the info I got isn't correct sorry for the scewed up info!

[turtle]

No worries, thanks for attempting to do the dirty work and get the scoop! :D