ATI R800 concept!? Wow!

[Wildside]

AMD: R8xx Speculation - Page 6 - Beyond3D Forum

http://i114.photobucket.com/albums/n273/Wirmish/ATi5870Concept.png

http://i114.photobucket.com/albums/n273/Wirmish/ATi5000Series.png

http://i114.photobucket.com/albums/n273/Wirmish/ATi5000SIMDCore.png

Concept -> 1 master chip and up to 4 slave chips.

The master chip is like a normal GPU... but with 4 SidePorts.
These SidePorts use the Rambus FlexIO Interface:
· Rambus FlexIO is capable of running from 400 MHz to 8 GHz.
· Contains only 12 lanes (5 lanes are inbound, 7 outbound)
· Theoretical peak I/O bandwidth of 76.8 GB @ 8 GHz (44.8GB out, 32GB in)
· Total bandwidth: 76.8GB/s x 5 SIMD Core (1 internal + 4 "slave chips") = 384GB/s

Q. Why use Rambus FlexIO Interface ?
A. Because AMD can use it (Rambus Signs Patent License Agreement with AMD) and because it's realy fast :
Quote:
"While Cell's XDR interface offers over 2x the memory bandwidth of any PC-based microprocsesor, Cell's FlexIO interface weighs in at 76.8GB/s - almost 10x the chip-to-chip bandwidth of AMD's Athlon 64." - Anandtech

A picture is worth a thousand words -> HD 5870 Concept
And this is the cards for the other market segments -> HD 5000 Series

A slave chip, or SIMD Core, is made of 8 blocs, for a total of 640 SP (128 SP 4D+1) and 16 TMUs -> picture

HD 5450 -> 640 SP + 16 TMUs (1 master chip)
HD 5650 -> 1280 SP + 32 TMUs (1 master chip + 1 slave)
HD 5670 -> 1920 SP + 48 TMUs (1 master chip + 2 slaves)
HD 5850 -> 2560 SP + 64 TMUs (1 master chip + 3 slaves)
HD 5870 -> 3200 SP + 80 TMUs (1 master chip + 4 slaves)

Good!:

· Need only 2 chips (master/slave) to create a complete series of graphic cards.
· The slave chip is an exact copy of the "SP/TMUs" part of the master chip.
So when you create the master chip, you also create 95% of the slave chip.
· Less time/money needed to create a new series -> 9 month product cycle instead of 12 months.
· Very small chips -> very good yields -> very good price.
· Shared memory architecture !

Bad!:

· The highest memory bandwidth hogs are the ROPs and the Level 2 texture caches, so ATI has physically placed these units next to the memory controllers and used a thousand or so traces to connect them. So, for now, it's impossible to put the RBEs inside the slave chips. Thus, the master chip has to contain enough RBEs for the HD 5870... but this high number of RBEs will be overkill for the HD 54xx/56xx. This is perhaps the major problem with this design.

NOTE: I know that there is a RBE in the first image but I'm too lazy to remove it.

i just found this, VERY COOL!!!!!!! I hope this is true!

 

[Baez]

Well it's most likely true but we'll have to wait to see if it really works.

 

[Apokalipse]

Who needs a 512-bit wide bus with this kind of bandwidth?

 

[zmatt]

wow! Simply wow! If that does come out 1H 2009 then Nvidia's GT212 wont hold a grain of hope against it. Of course that assumes that r800 works as they want it to. The fact that it is so modular will allow for cheaper cards. If anything ATI/AMD will have won the price war. Essentially they only have 2 different gpu types to make. The Master gpu, and then the slave, unlike in the past where each model up was physically a different chip.


Where did that guy get his information from? He didn't cite the website he pulled it from unless he made it himself. And if so then this isn't helping us much.

 

[Wildside]

wow! Simply wow! If that does come out 1H 2009 then Nvidia's GT212 wont hold a grain of hope against it. Of course that assumes that r800 works as they want it to. The fact that it is so modular will allow for cheaper cards. If anything ATI/AMD will have won the price war. Essentially they only have 2 different gpu types to make. The Master gpu, and then the slave, unlike in the past where each model up was physically a different chip.


Where did that guy get his information from? He didn't cite the website he pulled it from unless he made it himself. And if so then this isn't helping us much.

idk it just seems cool to read and look at.