At first glance, there is no serious cuts in computing power between GK110 and GeForce GTX Titan . Clock speeds have been increased even concerning the product Tesla. However, as explained colleagues, some restrictions still imposed. First, Titan does not support ECC, which formally should reduce the reliability of calculations. Second, HyperQ MPI and RDMA were disabled, needed to work in the computer clusters. However, if the GeForce GTX Titan is the only computing system accelerator.
Computing performance in single precision (FP32) 4.5 teraflops, with a double (FP64) - 1,3 teraflops. Earlier NVIDIA in their gaming solutions blocked the possibility to work with double precision at all, or reduces the number of active functional units responsible for these calculations. In the case of Titan is limited only by the clock frequency, because the use of the respective blocks provokes a rise in heat. The user can configure a driver to raise computing speed by doubles, but the frequency of the graphics processor can be automatically reduced to 725 MHz.
By default, the speed of the calculations in double precision is limited by the relation 1/24 FP32, to realize the full capacity, it increases to third FP32. Since the frequency of the processor is lowered from 837 to 725 MHz, we observe that the performance calculations of FP64 is not reduced by three times, as prescribed 1/3, but stronger, with 4.5 teraflops to 1.3 teraflops. Related Products :
|