Blazkowicz
Legend
I'm thinking he's playing funny games about the "GPRs". everyone would just say "registers" for a start.
Thanks, so it gives 16KB in registers for each 5D ALU, or rather 256KB per pack of 80 shaders. This is huge , this must be why the AMD architectures before GCN worked so well while looking quite more simple than the nvidia designs. AMD just seemed to pack blocks of SP togethers, eventually reaching 1600, while nvidia had their more complicated SM and floorplan already.
I've found a similar analysis for RV770 and it has the same 16K per 5D unit :
http://www.bit-tech.net/hardware/graphics/2008/09/02/ati-radeon-4850-4870-architecture-review/8
So, there was no difference between the original R700 series and its lone 40nm revision.
I quickly hunted for the 6870 GPU (Barts) then the 5870.
about the 5870 :
http://www.hardware.fr/articles/770-3/rv870-ou-cypress.html
about the 6870 : (it's all french and I explain below)
http://www.hardware.fr/articles/804-2/barts-tessellation.html
They say 6870 is basically a slightly cut down 5870 (a few less shader units and no FP64) with a few minor changes (texturing units, tesselation buffers, outputs).
I take this and say I'm pretty confident the register files didn't move.
So, at least every Radeon with a model number bigger or equal than 4350 and lower or equal than 6870 has the same amount of registers per set of units
Thanks, so it gives 16KB in registers for each 5D ALU, or rather 256KB per pack of 80 shaders. This is huge , this must be why the AMD architectures before GCN worked so well while looking quite more simple than the nvidia designs. AMD just seemed to pack blocks of SP togethers, eventually reaching 1600, while nvidia had their more complicated SM and floorplan already.
I've found a similar analysis for RV770 and it has the same 16K per 5D unit :
http://www.bit-tech.net/hardware/graphics/2008/09/02/ati-radeon-4850-4870-architecture-review/8
So, there was no difference between the original R700 series and its lone 40nm revision.
I quickly hunted for the 6870 GPU (Barts) then the 5870.
about the 5870 :
http://www.hardware.fr/articles/770-3/rv870-ou-cypress.html
bingo! we have 256KB per unit of 80 shaders.Chaque SIMD dispose de 16384 registres de 128 bits, comme les générations précédentes, de manière à supporter un nombre important de threads et ainsi masquer les latences autant que possible.
about the 6870 : (it's all french and I explain below)
http://www.hardware.fr/articles/804-2/barts-tessellation.html
They say 6870 is basically a slightly cut down 5870 (a few less shader units and no FP64) with a few minor changes (texturing units, tesselation buffers, outputs).
"Contrarily to Cayman, Bart's architecture remains globally unchanged from that of the Radeon HD 5000s"contrairement à Cayman, l’architecture de Barts reste globalement inchangée par rapport à celle des Radeon HD 5000
I take this and say I'm pretty confident the register files didn't move.
So, at least every Radeon with a model number bigger or equal than 4350 and lower or equal than 6870 has the same amount of registers per set of units