Cape VerdeCape Verde is a small GPU coming in at basically a third of Tahiti, used in the Radeon HD 7900s. Tahiti has a 384-bit memory bus and 32 fundamental blocks, the CUs, each of which contains 64 processing units and 4 texturing units. Cape Verde has just 10 CUs with a total of 640 processing units and 40 texturing units as well as a 128-bit memory bus.
This memory bus is fed by 16 ROPs charged with writing pixels to the memory. Tahiti has 32 ROPs and can output 2 triangles per cycle, which is no longer the case with Cape Verde. Here’s a visual of Cape Verde’s internal organisation:
Cape Verde adopts the specificities of the Graphics Core Next architecture that we described in detail here
. This includes an efficiency gain linked to the scalar type functioning of the processing units, a more advanced cache architecture, DirectX 11.1 support, PCI Express 3.0 support and various developments designed with GPU Computing in mind.
Cape Verde includes all the Tahiti specificities, with two exceptions:
- Double precision processing is at the equivalent of 1/16th of single precision against 1/4th on Tahiti, which makes sense on GPUs that haven’t been designed for professional processing.
- The L2 cache per 64-bit memory controller is up from 128 KB on Tahiti to 256 KB for Cape Verde, probably to prevent the total L2 cache from becoming too small.
Astonishingly, Cape Verde retains the 2 ACEs, which manage compute type tasks, as well as the 2 DMA engines loaded via the PCI Express port. It also supports PowerTune to control the card’s thermal envelope and ZeroPower, which almost turns off the GPU entirely in deep sleep mode (screen in standby).
Cape Verde and 2 of the 4 memory modules that supply it.
In spite of everything there are 1.5 billion transistors here. AMD has taken advantage of the 28nm fabrication process to fit them on a surface of more or less 125 mm², to be compared with 1 billion transistors engraved at 40nm that took up 172 mm² on Juniper, the GPU used on the Radeon HD 5700s and HD 6700s.