AMD's FireStream smashes teraflop barrier
One hot chip.
By Jeremy Kirk, IDG News Service | Published: 13:58, 16 June 2008
AMD has previewed the latest version of its high-performance chip packages, the FireStream 9250, which is due later this year.
Chips in the FireStream line offer much faster performance for mathematical calculations than other processors, the company claims. FireStream can take a single instruction and execute it using multiple sources of data in parallel.
The processor, which AMD refers to as a "stream processor" or General-Purpose General-Processing Unit (GP GPU) is capable of up to hundreds of parallel calculations per clock cycle, whereas other more general-purpose processors can only do a handful, the company said.
The 9250 will be available by the end of September for $1,999 (£1,000 approx). Last November, AMD released the FireStream 9170, the first such chip in the line that is capable of up to 500 gigaflops of computing power. A gigaflop is 1 billion floating-point operations per second. For single-precision calculations, the 9250 tops 1 teraflop, or 1 trillion floating-point operations per second, AMD said.
The FireStream chips can also perform "double precision" floating-point calculations. AMD says that helps runs intensive tasks such financial analyses and seismic processing much faster.
The 9250 comes with 1GB of GDDR3 RAM memory on its PCI card, a type of memory that was designed by ATI, the graphics chip maker now owned by AMD.
The 9250 will fit in a single PCI slot, making it compatible with most desktop systems, workstations and servers. Power consumption is less than 150 watts, similar to the 9170. Performance is rated at eight gigaflops per watt.
AMD has a software development kit, the AMD Stream SDK, to help developers create applications that use the full abilities of the FireStream chips. AMD says the FireStream line is not only for high-performance computing but also more mainstream and consumer applications.
