On Monday AMD announced the FireStream 9250, the company's latest general purpose graphics processing unit (GPGPU), and the successor to the FireStream 9170 product announced last year. Like the NVIDIA Tesla T10P also announced yesterday, the 9250 brings just over 1 teraflops of 32-bit (single precision) performance. The 9250 fits in a power envelope that the company says is less than 150 watts, and the card requires a single PCIe slot.

AMD focused hard on performance with this rev, not only getting to 1 teraflops of single precision performance and an 8 gigaflops per watt metric, but also outstripping the T10P by a factor of two in double precision performance. Combine this platform with AMD's commitment to making it easier for everybody to get those performance levels, and HPC users have a powerful option for bringing accelerated computing to an application near them.

Its friendly FLOPS per watt metric comes with some tradeoffs: the 9250 only sports 1 GB of GDDR3 RAM, four times less than NVIDIA's T10P, and half that on the original 9170. But unlike its competitor, the 9250 can handle an impressive 200 gigaflops in 64-bit computations, double the 9170's 100 gigaflops double precision FP performance. The new FireStream 9250 is scheduled to be available in the third quarter of this year. And with a $999 list price, it's substantially cheaper than the 9170 at $1,999.

While the timing of the announcement is clearly aimed at drawing interest from the HPC community, AMD is also focused on making this power available to a broader audience. The company is emphasizing steps aimed at making GPGPUs easier to program, and at making applications portable among different accelerators.

Last week, AMD announced partnerships with parallel tools providers Rogue Wave and RapidMind in an attempt to expand access to its hardware. The RapidMind partnership in particular is interesting in the context of this announcement since it announced support by RapidMind's Multi-Core Development Platform for the FireStream 9170 -- and one assumes that the 9250 can't be far behind. RapidMind's development tools allow one code base to be targeted at multicore processors, GPGPUs, and the Cell B/E without changing the code.

In a statement released with that announcement, Patricia Harrell, AMD's director of Stream Computing, said "AMD Stream embraces an open-systems approach to the GPU and is pleased to be working with RapidMind as we grow our ecosystem of software and service provider partners. RapidMind’s Multi-Core Development Platform gives our customers a unique advantage for harnessing the compute capacity of the AMD FireStream 9170 as well as multi-core CPUs."

AMD has also worked hard on creating effective interfaces for programmers to its GPGPUs, starting with the company's first accelerators in 2006 and the CTM, or Close to the Metal, programming interface. An early example of the use of acceleration and CTM at that time was the Folding@Home port, which saw a speedup of 30x over a CPU-only implementation. Programmer support continued to evolve with last year's announcement of the AMD Stream SDK, designed to help developers create accelerated applications for AMD FireStream, ATI FireGL and ATI Radeon GPUs. The SDK included AMD's implementation of the Stanford BrookGPU interface for stream computing, called Brook+, which is available as open source.

But this isn't an AMD-only world, and AMD recognizes that different programmer interfaces for different hardware is not the way to spur adoption of a technology over the long term. In an attempt to unify the programmer interface, AMD has announced its support for OpenCL.

The Open Computing Language is a community effort spearheaded by Apple (yes, makers of the iPod) to offer its customers a familiar, C-like environment for getting at the power in GPUs shipped with the company's Mac line of computers. OpenCL will be part of the technology rolled into Snow Leopard, the next release of Apple's Mac OS X operating system. In an interview in the NY Times Bits blog Steve Jobs said, "Basically it lets you use graphics processors to do computation. It's way beyond what NVIDIA or anyone else has, and it's really simple."

While OpenCL is still evolving, Patricia Harrell says that there are enough structural similarities between Brook+ and OpenCL that programmers shouldn't have a hard time making the transition. "OpenCL is a programming specification that feels like a C-level language, so it addresses the same level of abstraction as something like Brook+ or CUDA," says Harrell. "You can see both of those tools evolving to meet that programming specification and, while it's not yet complete, OpenCL already looks quite similar to what we already have."

While Apple has led the charge on OpenCL, the company has turned the effort over to the Khronos Group to manage as an open standard. In a statement released yesterday, Khronos confirmed its leadership of the effort and announced an impressive array of consumer and computational hardware partners, including AMD, ARM, Ericsson, IBM, Intel, Nokia, NVIDIA, and Samsung, among others.

Neil Trevett, president of the Khronos Group, said in Monday's statement that "this initiative is aimed at both desktop and embedded devices -- the day when you will be able to hold a supercomputer in the palm of your hand is perhaps not so far away."