Long before the official debut of the GeForce GTX Titan X appeared information that the graphics processor GM200 is much weaker in the calculations of double precision. By the way, the existence of Quadro M6000 can not be doubted. PNY representatives told our German colleagues that pin their hopes on the line Tesla accelerators update only with the release of the GPU Pascal, which is scheduled for 2016. Of this decision, we will tell more today in a separate article, but for now it is worth noting that in the segment Quadro memory capacity of 12 GB does not seem excessive to PNY company.
Meanwhile, colleagues from the site AnandTech thoroughly explain how NVIDIA managed to create quite compact GM200 graphics processor within the previous 28-nm process technology. The double-precision functional units in compact GM200 was reducexd a half times this allowed to increase the acceleration of graphics and computing single-precision functional units. As a result, transistor budget was able to dispose rationally: performance in graphics applications increased by 50% compared to GM204, and the core area increased by only 9% compared with the GK110 (Kepler). In GM200 NVIDIA managed to place 8 billion Transistors on an area of 601 square millimeters .
According to the head NVIDIA, video card GeForce GTX Titan X can be successfully used in the self-learning system of recognition of visual images. One of them, AlexNet, using GeForce GTX Titan X reaches the specified parameters for a little over two days, while the GeForce GTX Titan it takes six days, and the central Intel Xeon processor with 16 cores performs the task in 43 days. Moreover, combining in one system four video cards GeForce GTX Titan X, the period of training system can be reduced to 13 hours.
Ready computers DIGITS DevBox with four GeForce GTX Titan X company will sell for $ 15,000 to developers and researchers specialized applications. Related Products :
|