The HD 6770 (800 stream processors) is closest in performance to the HD 7750 (512 stream processors), with the HD 7750 coming out on top.
The HD 6850 (960 stream processors) is equivalent in performance to the HD 7770 (640 stream processors), with no clear winner.
The HD 6970 (1536 stream processors) is equivalent in performance to the HD 7850 (1024 stream processors), with no clear winner.
The current HD 7000 series cards have the following number of compute units:
HD 7750 = 8 CUs
HD 7770 = 10 CUs
HD 7850 = 16 CUs
HD 7870 = 20 CUs
HD 7950 = 28 CUs
HD 7970 = 32 CUs
From the above data, we can conclude that the HD 6670 would be equivalent to a 5 CU GCN part, the HD 7660D would be equivalent to a 4 CU GCN part, the performance of a 6 CU GCN part would fall somewhere between the HD 6670 and the HD 6770, and a 7 CU part would be a match for the HD 6770 with no clear winner.
It seems unlikely that the next gen APUs will have a GPU as powerful as the HD 7770, nor will Kaveri have one that is weaker than the HD 7660D and Kabini wont have one that is weaker than Brazos. In order for Kaveri to show a clear improvement across the board on Trinity in the GPU department, it would need 6 CUs. 2 CUs would be required to show an improvement upon the HD 6370D and a single CU would be enough to beat a HD 6320D. From this, we can conclude that the most likely configurations for the 28nm APUs will be the following:
HD 8550D = 1 CU
HD 8570D = 2 CUs
HD 8650D = 4 CUs
HD 8670D = 6 CUs
The HD 8550D will be used in Temash, the HD 8570D will be used in Kabini and the HD 8650D and the HD 8670D will be used in Kaveri. The 28nm integrated GCN GPUs will probably have specifications similar to the following:
HD 8550D (Temash)
400 MHz Engine Clock
1 Compute Unit (64 Stream Processors)
4 Texture Units
16 Z/Stencil ROP Units
4 Colour ROP Units
HD 8570D (Kabini)
600 MHz Engine Clock
2 Compute Units (128 Stream Processors)
8 Texture Units
16 Z/Stencil ROP Units
4 Colour ROP Units
HD 8650D (Kaveri)
800 MHz Engine Clock
4 Compute Units (256 Stream Processors)
16 Texture Units
32 Z/Stencil ROP Units
8 Colour ROP Units
HD 8670D (Kaveri)
900 MHz Engine Clock
6 Compute Units (384 Stream Processors)
24 Texture Units
32 Z/Stencil ROP Units
8 Colour ROP Units
Ignoring any improvements from the HSA features, optimisations to the GCN architecture, and taking into account the lack of GDDR5, the top end Kaveri will have GPU performance equal to or slightly better than a HD 6670. The faster the RAM the faster the performance and it will approach that of the HD 6770. The lower end Kaveri will have GPU performance similar tot Trinity. Temash would have slightly better performance than the HD 6320. Kabini on the other hand would show massive gains upon the HD 6320 and a fair improvement on the HD 6370. All of these advances will come with lower power consumption and take up less die space.
The HD 6850 (960 stream processors) is equivalent in performance to the HD 7770 (640 stream processors), with no clear winner.
The HD 6970 (1536 stream processors) is equivalent in performance to the HD 7850 (1024 stream processors), with no clear winner.
The current HD 7000 series cards have the following number of compute units:
HD 7750 = 8 CUs
HD 7770 = 10 CUs
HD 7850 = 16 CUs
HD 7870 = 20 CUs
HD 7950 = 28 CUs
HD 7970 = 32 CUs
From the above data, we can conclude that the HD 6670 would be equivalent to a 5 CU GCN part, the HD 7660D would be equivalent to a 4 CU GCN part, the performance of a 6 CU GCN part would fall somewhere between the HD 6670 and the HD 6770, and a 7 CU part would be a match for the HD 6770 with no clear winner.
It seems unlikely that the next gen APUs will have a GPU as powerful as the HD 7770, nor will Kaveri have one that is weaker than the HD 7660D and Kabini wont have one that is weaker than Brazos. In order for Kaveri to show a clear improvement across the board on Trinity in the GPU department, it would need 6 CUs. 2 CUs would be required to show an improvement upon the HD 6370D and a single CU would be enough to beat a HD 6320D. From this, we can conclude that the most likely configurations for the 28nm APUs will be the following:
HD 8550D = 1 CU
HD 8570D = 2 CUs
HD 8650D = 4 CUs
HD 8670D = 6 CUs
The HD 8550D will be used in Temash, the HD 8570D will be used in Kabini and the HD 8650D and the HD 8670D will be used in Kaveri. The 28nm integrated GCN GPUs will probably have specifications similar to the following:
HD 8550D (Temash)
400 MHz Engine Clock
1 Compute Unit (64 Stream Processors)
4 Texture Units
16 Z/Stencil ROP Units
4 Colour ROP Units
HD 8570D (Kabini)
600 MHz Engine Clock
2 Compute Units (128 Stream Processors)
8 Texture Units
16 Z/Stencil ROP Units
4 Colour ROP Units
HD 8650D (Kaveri)
800 MHz Engine Clock
4 Compute Units (256 Stream Processors)
16 Texture Units
32 Z/Stencil ROP Units
8 Colour ROP Units
HD 8670D (Kaveri)
900 MHz Engine Clock
6 Compute Units (384 Stream Processors)
24 Texture Units
32 Z/Stencil ROP Units
8 Colour ROP Units
Ignoring any improvements from the HSA features, optimisations to the GCN architecture, and taking into account the lack of GDDR5, the top end Kaveri will have GPU performance equal to or slightly better than a HD 6670. The faster the RAM the faster the performance and it will approach that of the HD 6770. The lower end Kaveri will have GPU performance similar tot Trinity. Temash would have slightly better performance than the HD 6320. Kabini on the other hand would show massive gains upon the HD 6320 and a fair improvement on the HD 6370. All of these advances will come with lower power consumption and take up less die space.