Mercury solution delivers sensor data directly to coprocessors for electronic warfare applications
NewsMay 11, 2012
CHELMSFORD, Mass. Engineers at Mercury Computer Systems, Inc., are leveraging technology from NVIDIA for a new product called StreamDirect, that is designed for sensor applications that require significant amounts of signal processing. StreamDirect essentially is an efficient way to deliver streams of sensor data directly to specialized coprocessors such as general purpose graphics processing units (GPGPUs).
The solution is geared toward electronic warfare programs then are heavily dependent on accurate sensor data, says Dinesh Jain, senior product manager, Mercury Computer Systems. The technology has already been deployed for a video imaging application, he adds.
The solution improves performance by enabling terabytes of raw sensor data to be processed in real-time via optimization of the transfers from the I/O sensors to the GPGPUs, supporting over 10 TFLOPS of processing capability in an OpenVPX system. StreamDirect was originally a custom solution designed for specific customer applications, that Mercury is now offering as a standard product -- the 6U OpenVPX GSC6201, Jain says.
It is designed to take advantage of the NVIDIA Fermi architecture and down the road the Kepler architecture, Jain says.
While many take advantage of NVIDIA technology, Mercury engineers say their solution is unique because it provides direct communication of data from the source, such as a sensor input component, into a coprocessor’s memory, likes a GPGPU, without intermediate storage in the CPU. In the past systems that used GPGPUs had to initially pass the data to a CPU’s memory and then transfer that same data from the CPU’s memory to the GPGPU. By eliminating this copy step, StreamDirect produces a direct, high-bandwidth DMA channel between the sensor and the GPGPU. The solution also uses Mercury’s POET/ICS technology and NVIDIA GPUDirect to enable system-wide communication capabilities for use in EO/IR, radar, cyber and electronic warfare applications.
StreamDirect enables GSC6201-based systems to have three times the performance of older generation systems by removing the intermediate data store-and-forward steps and allowing sensors to communicate directly with NVIDIA GPGPUs, says Scott Thieret, Mercury Computer Systems Technical Director. With this solution multiple GPGPU carrier cards can be hosted by a single CPU, which improves SWaP and GFLOPS/Watt, he adds.
For this solution Mercury engineers are using the MXM form factor in a rugged OpenVPX module. The GSC6201 is available today with NVIDIA’s embedded EXMF104 Fermi GPGPU and can be upgraded to the latest GPGPU architectures as soon as they are available.
NVIDIA CUDA, OpenCL and Mercury’s Scientific Algorithm Library (SAL) software development platforms are supported on the GSC6201. StreamDirect for NVIDIA GPGPUs is available today. The GSC6201 also is available today in commercial and rugged configurations including air- and conduction-cooled versions.
For more information on StreamDirect and the GSC6201, visit www.mc.com/gpgpu.
