ASEArch - Algorithms and Software for Emerging Architectures

FPGAs

FPGAs (field-programmable gate arrays) are a technology which has been around for the past 25 years, but is now starting to be used in some financial, engineering and scientific applications.

Maxeler is one of the leading companies in the exploitation of FPGAs, and they are providing a new FPGA dataflow supercomputer to STFC Daresbury.  In the past, developing applications for FPGAs has been very difficult, requiring developers with extensive computer science knowledge, but Maxeler is trying to address that challenge.

One of the key aspects of FPGA computing is that it is standard to use fixed point arithmetic, rather than floating point arithmetic.  The execution speed is also much faster if the requirements of the application allow you to perform the computation at very low accuracy.

For more information on the use of FPGA's, please see the talks from the workshop held at Daresbury in June, 2013.

 

Personal opinion (Mike Giles)

I arrived in Oxford in 1992, and at that time I was told that FPGAs were an emerging technology that would completely change high performance compuitng -- I am still hearing the same story today from its enthusiastic advocates.

The problem is that the application development effort is generally much larger than for any other technology, and new efforts to reduce it (for example through the adoption of OpenCL) may be at the expense of throwing away the performance benefits.

So, my view remains that it is a niche technology which is most suitable for integer-based applications, in areas such as bioinformatics, and perhaps for certain low-precision fixed-point arithmetic applications.

However, one big footnote to this comment is that in 2015 Intel acquired Altera (one of the two main FPGA companies) for almost $17bn, so Intel clearly sees a big future in FPGAs.