Nvidia GeForce GTX 680 Launch Review - PAGE 2

Like Fermi, Kepler GPUs are compiled of different configurations of Graphics Processing Clusters (GPCs), Streaming Multiprocessors (SMs), and memory controllers. The GeForce GTX 680 GPU consists of four such GPCs, eight next-generation Streaming Multiprocessors (SMX), and four memory controllers.

Starting at the top of the GK-104 block,  Kepler has a single GigaThread Engine which fetches the specified data from system memory and copies them to the frame-buffer. The Engine then creates and dispatches the threads from the memory to the GPCs, where it delivered to the execution unts. Following the GigaThread Engine are a total of four Graphics Processing Clusters (GPCs), which is where the majority of operations are performed. This is due to each GPC having a dedicated raster engine, as well as resources for shading, texturing and computation.

The memory sub-system of the Kepler architecture has also been redesigned to support higher speed clock speeds. This overhaul of the memory interface allowed Nvidia to push the operating frequency of the memory up to 6008MHz (4002MHz effective). The memory operates on a 256-bit wide GDDR5 interface, but since the good speed is at 6000MHz, the bandwidth has a 192GB/s rating. Additionally, the GK-104 GPU has four memory controllers, along with 512KB L2 cache, and since each GPC has its own Raster Unit there are a total of 32 Raster Operation Units.

Inside each GPC are two SMX units which have been optimized to offer the best performance-per-watt by running the the shaders at the same frequency as the GPU clock, and not double it. This approach gives Kepler twice the performance-per-watt of the Fermi architecture, while allowing more CUDA core to be packed into a single SMX unit. Inside each SMX are 192 CUDA cores which equates to a total of 1536 CUDA cores, triple the amount in the GTX 580. Of course since the CUDA core clock is equal to the GPU clock, the performance per CUDA core is reduced from the previous generation but the 1:1 clock design allows the GTX 680 to achieve the same throughout all while staying within a lower power envelope.

Looking at the functions of the execution units, the CUDA cores are designed to perform the pixel, vertex and geometry shading, as well as the physics compute calculations. The texture units on the other hand perform texture filtering, load/store units. fetch and save data to memory. Meanwhile Special Function Units (SFUs) handle transcendental and graphics interpolation instructions. Finally, the PolyMorph Engine handles vertex fetch, tessellation, viewport transform, attribute setup, and stream output.

The new Boost Clock feature is one of the biggest changes to the Kepler family. In essence the Boost Clock works along the same lines as Intel's Turbo Boost, which dynamically adjusts the clock speeds in real-time, thus increasing the performance. However, Boost Clock is different in the sense that the maximum Boost Clock frequencies are not necessarily where the GPU clock will cap during gaming. Instead, Boost Clock works at both a hardware and software level to dynamically boost the GPU clock speed and under most circumstances will increase the GPU clock speed well above the actual Boost Rating. Of course not all silicon is the same, so each Kepler board will have its own unique Boost Clock speed.

The typical board power defined for the GTX 680 is 170W. This means that the Boost Clock will increase the clock speeds to fit into this power envelope under load. Additionally, GPU Boost operates completely autonomously so there is no game profiles and no intervention required by the end user, providing an instant performance boost to gamers. The technology also works on a microsecond level, and does constant checks of the GPU voltage and conditions to see if the clocks can go higher or if they need to be throttled down to the base 3D clock.