512-bit memory interface with eight channels

GDDR3/4 = 32bit/chip = 32bit/channel
->
8 channels = 256bit

512bit @ 8 channels would require 64bits wide GDDR chips.
Chips like that don't exist.

http://www.samsung.com/Products/Semiconductor/common/product_list.a... (The chips on the pictures we have seen.)

Say hello to another..hmm...not right thing :P

Let me go out on a limp here. The 320 Steams is done with the same "math".

The confusion on memoy is that the gpu has a mem controller that handles mem in 64 bit chunks and uses 1 2 4 or 8 channels to handle it all.

The mem itself is either 32 or 16 bit per chip. Thus 512 bit has to use 8 channrls and 16 mrm chips while nvidia manafed to keep it down to 384 bit using 6 64 bit channels and 12 32 bit mem chips for high end abd the 320 bit 5 channel 10 chip version.
The odd thing is both nvidia and ati/amd seem tojump from high to 128 bit leaping past 256... very odd.

"HD" huh.....great...that's one less X at least.
Nice if it comes in at under $400.
A quick look at New Egg has 8800GTX selling from $530 to nearly $800.

The site also reports the R600 will have 320 stream processors, a 512-bit memory bus, and support for 24X antialiasing. By contrast, Nvidia's G80 chip has 128 stream processors, a 384-bit memory bus, and it supports up to 16X coverage sampled AA.

I'm interested in the way they implemented the extra AA samples, if true. Do they actually go for 4 samples per pass with an option for up to 6 loops or is it just some fancy way of adding together "temporal" samples, supersamples, and/or Crossfire options? (Anyways, improved AA would certainly be a positive surprise as far as I'm concerned.)

Also, as far as the processing cores are concerned, I suspect that "320 stream processors" is PRspeak for 4 blocks of 16 processors each capable of dealing with a Vec4 and a scalar per clock (i.e. 4/3 * Xenos), with the unavoidably worse utilisation rate, compared to the 128 scalar processors in G80. That could start a nice little DevRel war between AMD and nVidia, one lobbying for Vec4s, the other for Vec2s.

Other than that, I'm with flippy. Can we just please have some numbers and get it over with?

How many XXX's would we need to justify a $1000+ price tag?

3dmark06 isn't the best benchmark.

These all seem acceptable to me:

HD-2900
HD-2900-GT
HD-2900-XT
HD-2900-XL

Or even better:

HD-900
HD-900-GT
HD-900-XT
HD-900-XL

and then with each new generation change the prefix letters and maintain the numbers and postfix for segmentation purposes. This method of trying to find a higher and higher and higher and higher numbers and adding more and more and more Xs for each new GPU generation is getting out of hand.

Gah! Meant to reply to #4.

They use 8 64 bit channels and 16 32 bit chips.
The key factor tho is even with all this it supposedly isnt much netter then the nvidia card.

I own a Radeon HD 2900 XT and I'm really getting a kick out of these replies.

HD seems to be some internal techindustry moniker meaning DRM, first the intel 'HD' audio chip with its 'protected path' and now the r600 with its 'protected (for the MPAA) frame buffer' and 'onboard DRM soundchip'.

OHHH yay, more R600 rumors...

wake me up when they're in stores.

Never mind. (Sorry inkedsphynx, I changed my comment before I saw your post which noted my poor observational skills!)

320 stream processors....I haven't heard that number before???

So when is this card being released?

More hardware, less rumors.