Advanced Micro Devices is adding additional muscle to its line of FireStream graphics processing units with a new stream processor that looks to boost the company's offering around high-performance computing . AMD is also expanding its software development kit for developers creating applications for this growing field.
On Nov. 13, the same day AMD will release its latest quad-core Opteron processor called " Shanghai," the company will also launch the AMD FireStream 9270, a general purpose GPU (GP-GPU) that competes against the Nvidia Tesla 10 GPU in the HPC market.
The AMD FireStream GPUs are part of what the company refers to as ATI Stream, which looks to solve complex HPC problems by using the technology original development for the graphics market.
While HPC and supercomputers were based on more traditional CPU architectures for years, Nvidia and AMD are each beginning to develop new generations of graphics processors for the field that will allow applications to work faster and more efficiently. Unlike a CPU, a GPU or GP-GPU contains hundreds of smaller stream processing cores and allows the software's 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 AMD FireStream 9270 is one of several of these GP-GPUs that AMD has brought to the HPC market since November 2007. In June, the company offered the AMD FireStream 9250, which offers 1 teraflop or 1 trillion calculations per second of single-precision floating point performance.
With the FireStream 9270, AMD increased the single-precision performance to 1.2 teraflops and the company's engineers boosted the GPU's ability to handle double-precision performance to 240 gigaflops, which allows the FireStream 9270 to handle much more complex scientific applications and process data much faster.
The FireStream 9270 also offers up to 2GB of GDDR5 (graphics double data rate version 5) memory and works within a 160-watt thermal envelope. The AMD FireStream 9270, which will hit the market in December, sells for $1,499.
These types of GP-GPU for HPC are beginning to increase as all the major chip vendors are looking at new ways to increase performance and better handle the massive workloads needs in fields such as oil and gas exploration, mechanical design, finance and traditional science and research fields.
"Both oil and gas and finance have been two areas that have adopted this type of technology early," said Patricia Harrell, director of AMD's Stream Computing business. "In these areas, making a decision 50 milliseconds faster or a little bit better could be worth billions of dollars in these markets."
In addition to AMD and its FireStream GPUs, Nvidia offers its Tesla series of GPUs and the graphics company added to its lineup earlier this year with the Tesla 10 series that offers 1 teraflop of performance. Intel is expected to join the race in 2009 or 2010 when it releases "Larrabee," which is based on CPU cores instead of GPU cores.
While Nvidia is focusing on the GPU and Intel is sticking with the CPU technology it knows best, AMD is looking to bridge the gap and combine its traditional CPU technology with the graphics technology it inherited when the company bought ATI in 2006.
Eventually these two chip technologies will converge in what AMD is calling APUs, or accelerated processing units. The original name that AMD used referred to this as "Fusion." (An AMD spokeswoman at first said these APU processors will not appear until 2012 but the company called back later saying no firm launch date has ever been set.)
AMD is also supporting OpenCL, a type of computing language that allows the GPU to be programmed like a CPU and which is being supported by a number of companies including AMD, Intel, Nvidia, Apple and IBM. For its part, Nvidia has its own language called CUDA, or Compute Unified Device Architecture. ( AMD also support Microsoft's DirectX APIs.)
Rob Enderle, an analyst with the Enderle Group, suggested in an e-mail that for now, Nvidia has been edging out AMD thanks to both its GP-GPUs and its emphasis on CUDA as a programming language. Still, AMD and its ATI Stream processors and software have begun to put pressure on Nvidia.
"Overall this market is in its infancy, leadership at this market stage is often fleeting and both companies [Nvidia and AMD] would likely do better assuring things work on both platforms until this segment is more mature and then diverge," Enderle wrote in an e-mail.
"In other words, both would be well served to focus on growing the market not competing with each other," Enderle added. "For now most of these projects are one off anyway, but the need for inexpensive supercomputers is massive and the results, in terms of race benefits like curing diseases and solving energy problems, world changing. Given where this stuff plays, this may be one of the most important efforts from either company in their collective histories."
In addition to the new FireStream GPU, AMD is upgrading its software development kit for developers of HPC and scientific applications. The 1.3 version of the Stream Computing SDK is based on Brook+, a language variant of C and developed at Stanford University, and AMD has included improvements to the Brook+ runtime and kernel language to making application development easier.
In addition to FireStream and other developments concerning HPC, AMD also plans to release a new driver for its line of ATI Radeon HD 4000 series graphics cards that will allow consumers to take advantage of ATI Stream technologies. For example, in terms of applications, this driver update will allow a system that uses these discrete ATI Radeon graphics cards to convert high-definition video faster. This ATI Catalyst driver is set for a Dec. 10 release.