Packed full of ATI goodness, AMD is continuing on with its push to move graphics processors into the co-processor realm for common high performance computing applications.
The chip maker has pumped out it second generation stream processor. This time it's shifting focus from hardened developers to a broader HPC market — your friendly neighborhood medical imager, seismic modeler or computational fluid dynamicist.
The FireStream 9170 Stream Processor and accompanying development kit is a general purpose graphics processor unit, or GPGPU. Basically, it's a graphics chip modified to handle software that normally runs on mainstream server and desktop processors.
Most of today's general purpose processors handle computational tasks in succession. After one piece is done, the chip moves to the next. Parallel computing divides the tasks up and works on many pieces at the same time. The main goal of stream processing is to stage data so that it can be moved (streamed) through the memory system at high efficiency.
The catch is coding applications to run well for streaming, which can be a real bitch compared to standard x86 chips.
The 9170 has 500 gigaflops of compute power, which AMD theoretically places it in line with today's supercomputers. AMD says it will support double-precision floating point tailored for scientific and engineering calculations with the new product.
It consumes less that 150 watts of power, and can be added to existing servers or workstations via a PCI Express 2.0 x 16 interface. The processor board includes 2GB of GDDR3 (Graphic Double Data Rate 3) memory to relieve CPU traffic when handling large datasets. The circuitry is just 55 nanometers wide, smaller than the 65 nanometer circuits found in AMD's top-end computer processors.
Don't expect to be running Doom off this any time soon. Besides, you'd run a pretty fair risk of actually opening a portal to hell if you booted it up with this one.
The package is listed at $1,999, and is scheduled to be out during the Q1 of next year.
Nvidia has a similar product with its Tesla unit, and Intel is beavering away on an x86 chip to handle these HPC tasks.®