Seven years ago, during the announcement of the new console Sony - PlayStation 3 - was also presented the Cell processor . And presented in detail, including its per-block structure. PlayStation 4 is somewhat based on APU AMD's GCN generation graphics core and 8 cores Jaguar. We also know that the processor uses GDDR5 memory when simultaneously (or unified) treatment by the computing and graphics cores.
The new processor could get a block of internal memory eDRAM and 128-bit bus for treatment by the GPU to memory. This limited access to system memory at 88 GB / s, but could be compensated by the integrated power eDRAM. Instead, Sony decided to use the 256-bit bus with speed GDDR5 178 GB / s. This is done primarily with a view to simplify the lives of software developers, but in the production of APU with eDRAM promised to be easier than the APU with a 256-bit bus graphics core
Due to the limitation of opportunities for PCI Express 8 GB of computing cores and the graphics core can share a single pass about 1% of the memory. Unification of memory removes this limitation, as amended order of the GPU with the cache memory of the first and second levels. . All of this leads to delays of multiple layers.
Further, AMD has made three changes in the architecture of the APU. First, this kind of new bus for reading data from the graphics engine system memory directly, bypassing the L1 cache and L2. For a small amount of data that is subject to the rate of 20 GB / s, process synchronization and data caches are not needed, eliminating a lot of the processes for the exchange of information between the GPU and CPU.
Second, to run concurrent processes to process graphics and processing GPU non-graphic tasks in the data string to write to the L2 cache introduced a temporary marker - volatile bit, as it was called by Sony (po-angl). This marker simplified entry and returns data for individual processes, which does not lead to inhibition of graphics. In general - these are the 12 concurrent programs, which the developer wants.
Finally, Sony has asked AMD to increase the number of command sources for non-graphical calculations from two to 64 (original architecture GCN has a single source for the graphics commands and two for the computer). Six dozen sources allow to build an impressive turn of commands and flexibility to manage the queues.