Since it started in 1989, HOT CHIPS has been known as one of the semiconductor industry’s leading conferences on high-performance microprocessors and related integrated circuits. The conference is held once a year in August on the Stanford University campus in the center of Silicon Valley. The conference emphasis is to showcase real products and realizable technology, with this year’s presentation topics ranging from Intel’s next-generation processors to Xilinx and Altera’s latest FPGA platforms.
Maxeler Technologies’ Chairman, Michael J Flynn will describe Maxeler’s revolutionary MAXware technology at the HOT CHIPS 20 conference this August.
The standard approach to programming HPC applications is to use some language such as C++ with various augmentations implementing an application on as many multi core processors as possible, up to a point of node capacity or speed up saturation. This layered approach has some advantages, hiding implementation complexity to improve programmer productivity. This multi-layered approach limits acceleration or speedup. Large applications are partitioned over multiple such computational nodes. As a consequence such a node-parallel implementation suffers from bottlenecks in communication and limits computational efficiency (overall speedup/number of nodes) as well as scaling up cost, power, and cooling requirements. Such a highly parallel implementation increases programming, maintenance and porting complexity, in addition to raising reliability concerns.
MAXware is an alternative model which stresses maximizing the node speedup as far as possible before considering multi-node partitioning. MAXware uses an accelerator adjunct to the computational node, such as a large FPGA, GPU or any structured computational array. We achieve application acceleration in four stages: (1) Analysis, (2) Transformation, (3) Partitioning and (4) Implementation, during which an existing application is converted into a form where it makes best use of the available accelerator.
MAXware: acceleration in HPC shows how Maxeler’s combined methodology, tool set, accelerator hardware and software achieves significant improvements in price/performance compared to general purpose processors, with 1-2 orders of magnitude improvement in performance per watt and data center space requirements.