From HPCwire today we learn that DRC has announced the RPU110-L200, its new reconfigurable processing unit. This DRC solution integrates with the processor directly using AMD’s HyperTransport link through AMD’s Torrenza interface announced back in January. SRC announced a similar approach in June.
DRC’s RPU110–L200 module plugs directly into an open processor socket in a multi-way AMD Opteron system to provide direct access to adjacent double data-rate (DDR) memory and Opteron processors at HyperTransport speed and nanosecond latency. This tight coupling between the central processing unit (CPU) and memory eliminates bandwidth and latency issues and provides a general-purpose system with supercomputer capability by running computations in hardware.
DRC is making this announcement with Cray, which is planning to use the technology as part of their adaptive supercomputing thrust.
“DRC is working with Cray to deliver RPUs that offer efficient high performance computing alternatives to our customers. We believe that performance gains of 10 to 100X are possible on a number of applications when DRC’s RPUs are integrated into Cray’s systems,” says Jan Silverman, senior vice president, corporate strategy and business development for Cray. “This demonstrates yet another step forward in our Adaptive Supercomputing vision where multiple processing technologies are transparently available to the application.”
Update: This bit of interesting news from The Register
Sadly, the RPU (Reconfigurable Processor Unit) is only certified to work with a single Tyan motherboard. It’s a two-socket box based on AMD’s Rev E socket. DRC expects to have a Rev F product out by the fourth quarter. The new RPUs cost $9,500 each in volume (100 units) purchases.
Bad news indeed for hopes of wider adoption.