FPGA Gurus

Michael Feldman of HPCWire had an interesting piece last week about a financial supercomputing specialist, Kuberre Systems, that is using large banks of Altera Stratix-II FPGAs (four to 64, depending on the configuration), at the heart of what it calls HANSA, the Hardware Accelerator for Numerical and Systems Analysis). The role of HANSA in handling financial risk analysis and general-purpose pattern recognition led me to think about the changing role of HPC platforms, already evident at the last two years of the Supercomputing conference.

Vector and DSP specialists have long been aware of the turn to embedded functionality for algorithmic offload. Companies like Mercury Computer Systems specialize in single-board embedded functions for mil/aero and medical imaging. Integer operations have been more of a mixed bag, as companies such as IBM, Intel, Cray, and Convey still realize a significant business from standalone HPC platforms.

But the trend to embed algorithm acceleration in a former data-center world has been evident for many years. SRC Computers, the company formed after the demise of Seymour Cray’s second company, Cray Computers, has emphasized embedded applications. And if we examine the way similar HPC platforms have evolved, we find a trend to scaleable offload platforms for handling the sort of heavy-lifting tasks still best served by dedicated systems. But because of the way these systems are configured, FPGAs often end up being the most reasonable tool to consider.

When we say “embedded,” it doesn’t always mean that the HPC FPGA complex resides on a board that is directly placed in the data center server. It can also mean a computing complex in which the control plane and management OS is separated from the embedded bank of RISC and/or FPGA processors. In the case of Kuberre’s HANSA, a dual-core x86 is used for management under Linux or Windows A bank of 1 to 16 FPGA boards communicate with the control processor via an Algorithmic Processing and Switching Complex. The reconfigurable nature of the Stratix II (to be upgraded to Stratix IV) devices allows the FPGAs to be optimized for Eigenvalue, signals intelligence, crypto, and even bioinformatics searches.

This may represent a move by FPGA architectures to take over more embedded HPC functions in both integer and vector applications.