September 5th, 2012
NVIDIA Kepler GK106 GPU Detailed
Who wants to see some GPU pr0n? Before I began, let me explain that I made this diagram myself, but it’s based on quite solid knowledge about the GPU. So if you want to learn what does this new GPU add to the Kepler family – buckle up and read below.
NVIDIA Kepler GK106
NVIDIA’s GK106 is another iteration of 28nm Kepler GPU, its architecture is quite simple. This mid-range silicon will hold 960 CUDA cores in 5 Streaming Multiprocessors (SMXs). Those SMXs are divided into 3 Graphics Processing Clusters (GPCs), with one hiding only one SMX. Each SMX holds 192 CUDA cores and 16 Texture Mapping Units (TMUs). With a very sophisticated calculations we learn that the GPU has 960 CUDA cores and 80 TMUs. Further we can notice that the GPU has three Raster Engines. 192-bit memory interface is controlled by three 64-bit memory controllers. Each memory controller is tied to 128KB L2 Cache (the GPU has 384KB in total) and 8 Raster Operating Units (ROPs). The GK106 has 24 ROPs. The whole GPU is packed with 2.54 billion transistors (that’s exactly one billion less than GK104 — 29% decrease).
I have combined all the details below for your reading pleasure. What I don’t know yet is the exact size of the die.
| GK107* | GK106 | GK104** | GK110*** | |
|---|---|---|---|---|
| Transistors Count | 1.3 billion | 2.54 billion | 3.54 billion | 7.1 billion |
| Graphics Processing Clusters (GPCs) | 1 | 3 | 4 | - |
| Streaming Multiprocessors (SMXs) | 2 | 5 | 8 | 15 |
| CUDA Cores | 384 | 960 | 1536 | 2880 |
| Texture Mapping Units (TMUs) | 32 | 80 | 128 | 240 |
| Raster Operating Units (ROPs) | 16 | 24 | 32 | - |
| Texel Fill-rate | 13 Gigatexels/s | 78 Gigatexels/s | 128 Gigatexels/s | - |
| Memory Bandwidth | 28 GB/s | 144 GB/s | 192 GB/s | - |
| Memory Interface | 128-bit | 192-bit | 256-bit | 384-bit |
| Base Clock | 900 MHz | 980 MHz | 1006 MHz | - |
| Boost Clock | - | 1033 MHz | 1058 MHz | - |
| Effective Memory Clock | 1782 MHz | 6008 MHz | 6008 MHz | - |
* GK107 tied with GDDR3 memory
** Full GK104-400 GPU from GTX 680.
*** Some cards with GK110 will feature 13 and 14 SMXs, this is the full one from Tesla K20
For a reference here are GK104 and GK107.
NVIDIA Kepler GK104 / NVIDIA Kepler GK107
