As the evangelist for AMD Opteron™ processors, it isn’t often that I get to talk about some of the things taking shape on the consumer side of the housewith our Accelerated Processing Units (APUs). But every now and then something comes up that catches my attention.

For example, I recently blogged about running a home server on an APU. But that was a real low-power server with very light workloads; I chose the APU for the power efficiency, and the GPU side of my APU rarely ever comes to life unless I tunnel in via remote console for management.

But there is another side of the APU business that is quietly finding its way into servers these days, in high performance computing (HPC). In the HPC world, customers are clustering large numbers of powerful servers together to help solve large, complex problems. Traditionally, these have been dual processor servers, but when it comes to creating the capacity for a large number of floating point operations per second, customers have increasingly been turning to GPUs to deliver that performance.

AMD FireStream™GPU compute accelerators bring the high FLOP count to HPC computing, but others are finding more creative ways to achieve high FLOP count by leveraging the GPU portion of the APU. Penguin Computing, one of the earliest and most vocal users of AMD Opteron technology, has brought a new server to the market: the Penguin 2A00 server. This server is built around an AMD APU and is the first of its kind that has been deployed on a larger scale in an HPC environment.

What makes APUs interesting in the HPC world is that, in contrast to traditional CPU-GPU hybrid compute models, CPU and GPU share a common physical memory space. Latency due to data transfers between main and GPU memory is avoided. Data does not have to be duplicated and GPU programmers no longer have to think about data locality in terms of GPU and main memory, while still using the familiar OpenCL framework to program their applications.

Sandia National Labs, another strong proponent of AMD Opteron technology, sees this new product as a great platform for their research into future programming models. Sandia is particularly interested in the combination of low latency interconnects that support RDMA (Remote Direct Memory Access), such as Infiniband, and APUs. A solution that brings these technologies together could be a great platform for a new generation of distributed GPU applications.

As we approach Supercomputing 2011 and the launch of our processors codenamed “Interlagos”, it is clear that the HPC landscape is changing and that the traditional paradigms of supercomputing are continuing to evolve. Just a decade ago, most of the world’s supercomputers were built on larger, proprietary platforms. While these delivered high performance, they were much more expensive, difficult to program, and were not able to keep pace with the rapid levels of innovation that the HPC community demanded.

X86-based servers became much more important to the HPC landscape, and 64-bit memory addressing helped them really rise to prominence in the recent past. Now GPU technology is helping to redefine the HPC space again, and clearly AMD Fusion APUs will have a place in that computing landscape. With both leading CPU and GPU technology, AMD has the technology pieces to assemble a wide range of solutions for HPC, as well as regular server workloads. One day we may even see APUs become a staple in the HPC space, just as our CPUs and GPUs have helped to solve customer problems and bring new levels of performance. Clearly, there is an interesting future in high performance computing, and this year’s Supercomputing conferencewill showcase some of the more creative solutions.

Stop by and say hello. We’ll be at booth #823.

John Fruehe is the Director of Product Marketing for Server, Embedded and FireStream products at AMD. His postings are his own opinions and may not represent AMD’s positions, strategies or opinions. Links to third party sites, and references to third party trademarks, are provided for convenience and illustrative purposes only. Unless explicitly stated, AMD is not responsible for the contents of such links, and no third party endorsement of AMD or any of its products is implied.