by Anton Shilov
09/21/2010 | 04:53 PM
On Tuesday for the very first time in the history of Nvidia Corp. as well as graphics processors industry in general the company shared at its GPU Technology Conference the names and some of the features of its forthcoming graphics processor architectures.
On the first day of the GPU Technology Conference (GTC), Nvidia has unveiled brief details about its next-generation graphics processing units (GPUs). Quite logically, the future processor architectures by the company - code-named Kepler and Maxwell - will be massively focused on general-purpose computing and will offer substantial performance increases compared to existing chips.
Next year - sometimes in the second half of 2011 - Nvidia plans to release GPUs based on the Kepler architecture. The chips will be made using 28nm process technology and will bring tangible performance improvements compared to currently shipping graphics chips based on Fermi architecture. According to Nvidia, Fermi architecture is capable of achieving typical double precision (DP) performance of 1.5GFLOPS per watt. Kepler architecture will increase performance per watt by about 3 or 4 times, hence, it is possible to expect something like 1.125TFLOPS - 1.50TFLOPS of DP performance by a chip with 250W thermal design power. Although math performance with DP accuracy does not have a direct correlation with graphics performance it is possible to expect Kepler to be at least two or three times faster than Fermi in games.
Nvidia Maxwell will be launched in 2013, according to chief executive of Nvidia Jen-Hsun Huang. Given the timeframe, it is logical to expect 20nm process technology to be used for manufacturing of Maxwell. The architecture due in three years from now will offer whopping 14 - 16GFLOPS of DP performance per watt, a massive improvement over current-generation hardware.
"Between now and Maxwell, we will introduce virtual memory, pre-emption, enhance the ability of GPU to autonomously process, so that it's non-blocking of the CPU, not waiting for the CPU, relies less on the transfer overheads that we see today. These will take GPU computing to the next level, along with a very large speed up in performance," said Jen-Hsun Huang.
It is noteworthy that Nvidia has decided to refresh the GPU architectures every two years, not in twelve - eighteen months as it used to do before and its rival ATI, graphics business unit of Advanced Micro Devices, does nowadays. It looks like Nvidia has faced rather massive limitations with process technologies from contract makers of semiconductors, which seems to slowdown the company's business. Still, even availability of process technologies does not limit Nvidia's abilities to refresh its actual products lineups to stay competitive.