Korea develops new software for world’s fastest supercomputer
A team of Korean scientists has developed a new software that uses graphic processing units (GPUs) instead of central processing units (CPUs) as a main computing device to build the world's fastest supercomputer, the science ministry said Tuesday.
The use of GPUs instead of CPUs for general computing is already common for supercomputers, according to the Ministry of Education, Science and Technology. Most existing supercomputers are built on the mechanism known as "general purpose computing on GPU," but currently use only two GPUs per node.
The new mechanism or software, developed by the team from Seoul National University, connects at least three and up to six GPUs to each node or connection point, making its end-product much faster and more efficient.
The team, led by Lee Jae-jin, a professor at the university, used the technology to build its own super computer, SnuCore, that has 16 nodes with each node equipped with two CPUs and six GPUs.
In a speed test, known as the Linpack benchmark, the SnuCore recorded 0.991 TFLOPS (tera-floating-point operations per second), the highest among the world's top super computers, with its efficiency, in terms of MFLOPS per node, ranked the 20th.
"This (development of the software) will be an opportunity for the country to finally catch up with other leading countries in supercomputers," Lee was quoted as saying. (Yonhap)