14 DAY TRIAL // Many leaks, rumors and other sorts of reports on NVIDIA's upcoming GPU series are roaming the Internet, but the latest one may be somewhat more relevant than the rest.
The folks at TechPowerUp claim that some “reliable sources” provided them with previously unknown details about the NVIDIA Kepler GPU, the graphics processing unit that will power the next-generation collection of graphics adapters, known as GeForce 600 series.
What the new report says is that, while the basic relationships between the GPU's components will be relatively unchanged compared to the Fermi, the same cannot be said about the CUDA side.
Instead, the Santa Clara, California-based company chose to arrange everything in such a way that a greater speed was achieved during parallel processing.
Thus, in the case of the GeForce GK104, the TMU / Geometry Domain involves 32 ROPs (Raster operation Units) and 8 TMUs (texture units) per SM (Streaming Multiprocessors), leading to a total of 128 TMUs.
Speaking of the SMs, they are 16 in number, divided into four Graphics Processing Clusters (GPCs). Given that there are 96 Stream Processors (SPs) per SM, that leads to 1,536 CUDA cores (triple compared to Fermi GTX 580).
Moving on, there are 2 GB of memory (2,048 MB), which function on an interface of 256 bits.
As for everything else, the GPU works at 950 MHz and, as one of the most important changes, there are no hot-clocks for the CUDA cores.
Previously, CUDA cores worked at a higher frequency than the GPU, but from now on they will stick to the same number of MHz.
Finally, the 2 GB of VRAM boast an actual clock of 1,250 MHz (5.0 GHz effective), enabling a bandwidth of 160 GB/s. Everything is crammed inside a chip with a die area of 340mm².
All in all, the Kepler GPU is expected to deliver 2.9 TFLOP/s single-precision floating point compute power and 486 GFLOP/s double-precision floating point compute power.