ArchitectureThe GeForce GTX Titan is based on NVIDIA's biggest chip for the "Kepler" micro-architecture, codenamed "GK110." First introduced as part of the Tesla K20 GPU compute accelerator, the chip is built on the 28 nanometer silicon fabrication process and packs a staggering 7.1 billion transistors. That's over three times the transistor count of Intel's 8-core Xeon "Sandy Bridge-EP" processor. Its component hierarchy is identical to that of other GPUs based on the architecture.
While the GK104 features four graphics processing clusters (GPCs) with two streaming multiprocessors (SMXs) each, the GK110 features five GPCs with three SMXs each. The SMX design hasn't changed: it still houses 192 CUDA cores each, so the physical CUDA core count on the chip works out to 2,880. The GeForce GTX Titan feature-set only includes 2,688, since one of the 15 SMXs is disabled. This probably helps NVIDIA harvest the GK110 wafers better by giving TSMC room to get one SMX "wrong."
The GK110 features a total of 240 texture-mapping units (TMUs), but since TMUs are contained within SMXs, the GTX Titan ends up with 224. Since raster operations processors (ROPs) are tied to the memory bus width on the Kepler family of GPUs in general, the GK110 features 48 as compared to the 32 on the GK104. Speaking of memory, the GK110 features a 384-bit wide GDDR5 memory interface. To our surprise, NVIDIA made 6 GB the standard memory amount. While no game needs that much memory, even at 2560 x 1600, NVIDIA is probably looking to stabilize 3D Vision Surround performance. The GTX Titan lets you set up a 3D Vision Surround setup with three 2560 x 1600-pixel displays.
In terms of clock speeds, the GeForce GTX Titan core doesn't get anywhere close to the 1 GHz mark, which is characteristic of every big chip from NVIDIA. The core is clocked at 838 MHz, with a nominal GPU Boost frequency of 878 MHz, while the memory is clocked at 6008 MHz, so the card ends up with a memory bandwidth of 288 GB/s.
With the GeForce GTX Titan, NVIDIA also introduced its second generation GPU Boost technology. The technology allows GPUs to increase their core clock speed and voltage beyond nominal values while respecting a set power target. The new GPU Boost 2.0 takes temperatures and additional power draw into account when adjusting core clock speeds and supportive voltages. This ensures that even stressful applications get the benefit of higher clock speeds so long as the GPU isn't overheating, which should particularly please enthusiasts who overclock their cards using such sub-zero cooling methods as liquid nitrogen or dry-ice evaporators.