AMD: R8xx Speculation
- Thread starter Shtal
- Start date
From top to bottom:
Hemlock(R800), Cypress(RV870 300mm²), Cedar(RV840/~225mm²), Juniper(RV830/180mm²) and Redwood(RV810/120mm²).
If the line-up is wrong, I'll change my signature accordingly, for now. The specs that guy gave over at pconline were regurgitated, much like all the posts. The info came Endgadget/Inq where they show a Cypress with a launch/event date in September on the "17Mile Drive" course at Pebble Beach
Hemlock(R800), Cypress(RV870 300mm²), Cedar(RV840/~225mm²), Juniper(RV830/180mm²) and Redwood(RV810/120mm²).
If the line-up is wrong, I'll change my signature accordingly, for now. The specs that guy gave over at pconline were regurgitated, much like all the posts. The info came Endgadget/Inq where they show a Cypress with a launch/event date in September on the "17Mile Drive" course at Pebble Beach
Last edited by a moderator:
Are you implying that it'll be a rainmaker or that it'll make baby Jesus cry?
Yes. 
So the whole RV830 being cancelled and larger than RV840 is wrong?
LMAO. Awesome answer.Yes.
Jawed
Legend
Maybe the 64bit segment is dead. Nobody buys 64bit GPUs for €32 if a 128bit GPU, which is more than twice as fast, costs €40.
The other possibility is, that 128bit (G)DDR3 part is cheaper to manufacture than a 64bit GDDR5 part. GDDR5 GPUs require some more die-space for the exatra pads - minimum die-size for 64bit GDDR5 interface could be close to minimum die-size for 128bit (G)DDR3 interface.
The other possibility is 128bit (G)DDR3 interface at 40nm now and 64bit GDDR5 after die-shrink.
Anyway, the 180mm²/300mm² die-size difference is identical (percentually) to RV730/RV770. Could it mean, that the 300mm² part carries 2,5x SPs (800->2000), just like RV730/RV770 (320->800)?
The other possibility is, that 128bit (G)DDR3 part is cheaper to manufacture than a 64bit GDDR5 part. GDDR5 GPUs require some more die-space for the exatra pads - minimum die-size for 64bit GDDR5 interface could be close to minimum die-size for 128bit (G)DDR3 interface.
The other possibility is 128bit (G)DDR3 interface at 40nm now and 64bit GDDR5 after die-shrink.
Anyway, the 180mm²/300mm² die-size difference is identical (percentually) to RV730/RV770. Could it mean, that the 300mm² part carries 2,5x SPs (800->2000), just like RV730/RV770 (320->800)?
Jawed
Legend
AMD can only sell non-IGP graphics for Intel CPU systems, so having no "RV810" would lock them out entirely from the next rung up from IGP. I freely admit, I've no idea of the sales for RV710s as stand-alone boards or in pre-built systems. Or what ATI+NVidia sales for this rung are in Intel systems.
Or, whether D3D11 is, in itself, relevant.
And, ahem, what about "professional" cards based on 64-bit? No idea what happens already...
As to 2000 ALU lanes, 300mm² sounds like a budget was bust. If the trend continues, how long before a major GPU is ~400mm²?
Jawed
Or, whether D3D11 is, in itself, relevant.
And, ahem, what about "professional" cards based on 64-bit? No idea what happens already...
As to 2000 ALU lanes, 300mm² sounds like a budget was bust. If the trend continues, how long before a major GPU is ~400mm²?
Jawed
OpenGL guy
Veteran
RV515 supported 64- and 32-bit operation and there were such products. Previously, 64-bit was the smallest memory configuration allowed.
I had an X300 SE that was most definitely only 64-bit.
Jawed
Legend
I'm thinking that AMD would want a 64-bit chip with an area of ~60mm². Being able to produce such a chip in a timely fashion seems like part and parcel of the sweet-spot strategy - and is presumably extra important on a D3D-version-inflection.
It would be nice if turning on tessellation with even the slowest 64-bit D3D11 GPU would increase performance, as tessellation, in general, should be used as a form of geometry compression...
Jawed
It would be nice if turning on tessellation with even the slowest 64-bit D3D11 GPU would increase performance, as tessellation, in general, should be used as a form of geometry compression...
Jawed
OpenGL guy
Veteran
The ASIC was 128-bit, but the X300 SE boards only had half the memory channels populated with memory from what I recall.
Yes there were many products with some disabled functionality, but I'm speaking about native bus width. RV370 and RV515 were 128bit, while RV610 and RV710 were 64bit.
Anyway, you're right... RV810 could be 128bit GPU, but value models can have half of the memory bus disabled...
Was this already mentioned?
http://forum.donanimhaber.com/m_31456023/mpage_1/key_//tm.htm#31456023
Basicly it seems like summing up the earliest rumored specs, though
http://forum.donanimhaber.com/m_31456023/mpage_1/key_//tm.htm#31456023
Basicly it seems like summing up the earliest rumored specs, though
That link claims 1.5x the ALUs, 2.4x the TMUs, and 2x the ROPs (all with ~1.1x the bandwidth). I haven't been following things too closely, but are we really expecting a lower ALU:TMU ratio?
Looks reasonable to me. RV7xx generation was mostly TMU limited, not ALU limited, so increasing TMU's will give biggest performance improvement.
And bandwidth is expensive.
And they have improved the schedulers a lot, so it may be that they get better efficiency from their shader processors, so effective real-world shader performance may increase more than 1.5 times.
Similar threads
