A Fresh Look at RPUs

Share it on Twitter  
Share it on Facebook  
Share it on Linked in  

The ever-growing demand for increased performance and decreased power consumption in the data center is causing many researchers to take another look at reconfigurable processing, particularly in the burgeoning field of high-performance computing.


Although approaches may vary, a reconfigurable computer generally consists of a standard CPU linked to a series of field-programmable gate arrays (FPGAs). Because the FPGA can be reconfigured on the fly, this reconfigurable processing unit (RPU) can be configured to perform a specific task at high speed, and then reconfigured for the next task.


How fast are they? Some people are talking about a 1,000-fold increase in performance, according to HPCwire, albeit for extreme HPC functions like gene sequencing. But in general, because the FPU offers hardware-based parallel processing and does away with time-consuming operations like instruction fetching, performance boosts of 100 percent are not unreasonable. This, of course, comes at the expense of flexibility and programming ease that typical CPUs provide.


Commercial development is already taking place. SRC Computers recently announced its support for AMD's Torrenza initiative, which will bring SRC's Implicit+Explicit architecture and Carte programming environment to Opteron-based servers, primarily the HP ProLiant DL 385. SRC is looking to release an AMD system capable of supporting ANSI Standard C and Fortran environments by the end of the year.


Another firm, DRC Computer, says it has an FPGA-based coprocessor that plugs right into a standard socket on an Opteron motherboard running Linux. By connecting directly to the main HyperTransport processor, the device is said to eliminate bandwidth constraints and bottlenecks that normally come with coprocessors on HPCs. A single HT connection gets you 3.2 Gbps, although each RPU can be configured for three HT connections, delivering bandwidth up to 9.6 Gbps.


Reconfigurable computing is also making a mark in very large-scale integration (VLSI) because it can be made to operate like just about any circuit the designer can imagine. Improved hardware support, enhanced microprocessor optimization, and the ability to quickly design prototypes of all sorts are just some of the possibilities.


Reprogrammable computers are unlikely to sweep the tried and true CPU off the map completely. But for specialized applications and data-intensive environments like life sciences and financial services, they could prove to be a valuable tool in the green data center.