Advanced Micro Devices is looking to further its reputation within the high-performance computing field with a new graphics chip that looks to take better advantage of parallel computing.
The chip maker Nov. 8 announced the FireStream 9170 GPU (graphics processing unit), along with a software development kit, which AMD plans to bring to the HPC market in the first quarter of 2008. The new processors cost $1,999, according to AMD.
In the traditional method of computing, a CPU funnels all the data through a single path and gains performance through greater and greater clock speeds. However, a GPU allows the softwares instructional threads to run in parallel, breaking the information down into smaller pieces, which provides for high throughput and better performance for various applications without relying as much on cranking the clock speeds.
The FireStream chip is designed for industries that need to crunch massive amounts of data, such as various scientific fields, oil and gas, and engineering.
AMD developed the FireStream chips through the technology it acquired from ATI in October 2006. Besides AMD, Nvidia has also begun developing GPUs as a viable option for HPC. In June, Nvidia unveiled its Tesla GPU, which will target the same types of HPC applications that FireStream will when it comes to market next year.
Before the sale to AMD, ATI announced that it would go ahead with a project known as “Stream Computing,” or GP-GPU (general-purpose computing on GPUs). This project, which AMD has taken up under its own name, looked to use the massively parallel processing power of graphics cards to perform a number of nongraphics tasks, such as crunching numbers of a spreadsheet.
In addition to these types of applications, AMD is using ATI technology to work toward combining the CPU and GPU on the same piece of silicon under a project codenamed “Fusion.” While FireStream is clearly meant for HPC and supercomputers, the results of Fusion will likely target more mainstream desktops PCs and notebooks.
The FireStream GPU is built on a 55-nanometer manufacturing process—most of AMDs mainstream processors, such as the Opteron, are built on a 65-nm process—and will consume less than 150 watts of power. The chip offers 2GB of memory and a peak performance of 500 gigaflops, or 500 billion floating point calculations per second. Nvidias Tesla GPU, by comparison, offers 518 gigaflops of performance.
AMD will offer additional details about FireStream at the SuperComputing expo, which starts Nov. 12. In addition to the new GPU, AMD is offering a new software development kit for ISVs that gives developers access to certain APIs and other specifications, enabling them to take advantage of FireStreams capabilities.