by Anton Shilov
02/08/2010 | 07:21 PM
The talks about graphics processors powering servers have been around for the last three years, but so far only a number of special-purpose supercomputers take advantage of graphics processing units (GPUs) and their extreme amount of cores. But in two years time the situation may become different, claims Advanced Micro Devices and GPUs may find themselves even inside mainstream servers
“Mainstream servers in the future could have a combination of graphics processors and CPUs in servers as applications take advantage of thousands of GPU cores”, said Gina Longoria, director of the product management and workstation division at AMD, reports IDG News Service. According to Ms. Longoria, servers may get GPUs as soon as in the next two years.
In the next two years several rather important introductions will take place: in 2011 AMD plans to introduce its highly-anticipated Bulldozer micro-architecture along with 16-core microprocessors, and in 2010 – 2011 timeframe the architecture of graphics processors will become more general and such chips will be more suitable for general-purpose computing. Even now graphics chips can provide formidable double-precision floating point performance, however, their programmability is not as flexible as that of central processing units (CPUs), therefore, the number of applications that can take advantage of GPGPU technologies is limited nowadays.
“As GPU becomes more relevant, that's a better way of getting performance than [CPU] cores,” Ms. Longoria said.
It is not clear how AMD plans to squeeze graphics processors into servers. Back in the days AMD already introduced Torrenza initiative under which special-purpose accelerators could be installed into the same sockets as AMD Opteron processors in multi-processor servers in order to gain performance in specific applications. It remains to be seen whether AMD will resurrect Torrenza or will simply propose to install forthcoming FireStream accelerators into PCI Express slots. In the latter case the solution will hardly be adopted by large data centers due to size, complexity and power consumption. On the other hand, installing GPUs into processor servers will reduce their horsepower due to lower speed of memory that is installed into slots and is not directly connected to GPUs, like on graphics cards.