|
Shortly before the announcement of the first graphics solution based on Kepler architecture , we have had the opportunity to study the schematic structure of GK104 processor and estimate the scale of changes that NVIDIA has made since Fermi architecture . First Kepler received four blocks Graphics Processing Cluster (GPC), which contains eight multiprocessors SMX 1536 CUDA cores .
Younger decision - GeForce GT 640 GK107 processor - has only one GPC and, accordingly, two multi-processor with 384 cores CUDA. the scheme is as follows:
Associate published the schema of GK106 processor. Interestingly, one of the three blocks GPC solution consists of a single multiprocessor. Thus, in the circuit are 5 SMX and 960 cores CUDA. Other elements that somehow affects GPU performance, you can also count, just remember part of a multiprocessor:
- block geometry PolyMorph 2.0;
- first cache level (64 KB), cache and cache textures instruction;
- 4 scheduling of data flow Warp Scheduler;
- 8 instructions dispatchers;
- file register128 Kbytes;
- 192 core CUDA;
- 32 units special functions SFU;
- 32 units load and save data LSU;
- 16 blocks TMU.
The cluster does not include the power distribution GigaThread Engine, second cache level, bitmap blocks (a total of 24 ROP) and three blocks of memory controller:
So far we know only one product on base of GK106 processor - this is the GeForce GTX 660. However, based on this GPU will be released, at least four products.
GK106-400 will be known as the GeForce GTX 660 with 960 cores CUDA, 2 GB memory and 192-bit bus access. Modest GK106-250 (maybe, GeForce GTX 660 SE) - one multiprocessor is likely to disabled and remain 768 cores CUDA. Finally, GK106-200 version will be the base for the GeForce GTX 650 Ti with 576 CUDA cores and 128-bit memory interface. Related Products :
|