SCIENCE
Curtiss-Wright Controls Introduces Industry's First NVIDIA GPGPU OpenVPX Engine Based on NVIDIA Fermi Architecture
Curtiss-Wright Controls Embedded Computing (CWCEC) has announced its first general purpose graphics processing unit (GPGPU)- multi-core engine, the VPX6-490 GPU Application Accelerator. The VPX6-490 features dual NVIDIA GPUs based on the NVIDIA Fermi architecture, each with 240 CUDA cores. Integrated into a High Performance Embedded Computing (HPEC) subsystem, VPX6-490 functions as a co-processor attached to a host Intel-processor board and takes advantage of the new PCIe Expansion Plane definitions in the VITA 65 OpenVPX standard to provide off the shelf backplane support for high-speed interconnection between pairs of SBC/GPU. The combination of 2nd Generation Intel processors, Gen2 PCI Express interconnect and 480 NVIDIA CUDA cores raises the performance bar for embedded High Performance Embedded Computing (HPEC) systems for demanding military digital signal processing (DSP) applications such as C4ISR, EO/IR, and SatCom.
"The VPX6-490 uniquely brings NVIDIA's new 240 core graphics processors to the rugged deployed environment, providing system integrators with the highest performance GPGPU embedded building blocks available," said Lynn Bamford, vice president and general manager of Curtiss-Wright Controls Embedded Computing. "Combined with our OpenVPX single board computers, FPGA engines and FMC I/O modules, we now offer the industry's most comprehensive and highest performance GPGPU solution for the most demanding HPEC applications."
"Rugged open-architecture processing modules, like Curtiss-Wright's VPX-490, will enable military system designers to leverage the massively parallel performance of our newest GPUs," said Taner Ozcelik, general manager, embedded and automotive business, NVIDIA.
The VPX6-490 takes full advantage of GPUs based on the NVIDIA Fermi architecture. Designed for high performance computing, the newest generation of NVIDIA processors feature larger internal shared memories, a completely new L2 cache, unified memory addressing and many other enhancements to improve CUDA-based applications performance and improve programmer productivity. The VPX6-490 supports its dual high performance GPU processors with a 2 Gbyte, 256-bit wide, 80 Gbyte/s GDDR5 memory subsystem designed to eliminate data bottlenecks and support large signal processing datasets into the onboard memory. Each GPU device supports a full 16-lane Gen2 PCIe interface to the backplane, supporting the maximum possible bandwidth between host and GPU. The VPX6-490 also supports 8-lane and 4-lane PCIe interfaces.
The VPX6-490 is designed to work with Intel processor- based SBCs, such as the Curtiss-Wright VPX6-1956 2nd Generation Intel Core i7 and the CHAMP-AV8 dual 2nd Generation Intel Core i7 multi-processor card.
VPX6-490 Performance Features:
- 6U OpenVPX GPGPU Application Accelerator
- 80 Gbyte/s memory bandwidth
- Enhanced GPGPU performance based on NVIDIA Fermi architecture:
- 240 CUDA processor cores
- 2 Gbytes GDDR5 memory
- 256-bit memory interface
- 16-lane Gen2 PCIe interface
- Dual DVI graphics outputs
- CUDA 4.0 library support
- Rugged air-cooled and conduction-cooled variants
- Temperature sensors
- IPMI
- 12V power
Curtiss-Wright Continuum HPEC Subsystem Components
The VPX6-490 is a key element of Curtiss-Wright's new Continuum High Performance Embedded Computing (HPEC) initiative. Continuum HPEC systems consist of a large number of distributed processors, IO, and software stacks connected by a low latency system fabric. HPEC capabilities are developed in our Ashburn, VA HPEC Center of Excellence. With scalable architectures, dataflow modeling and configuration validation, Curtiss-Wright's Continuum HPEC customers can source embedded supercomputing platforms that integrate Intel-based multi-processor boards with AVX, GPGPU co-processors, Xilinx Virtex 6 FPGAs, SRIO and Ethernet switching with Open Standard software solutions including VxWorks, and Linux with OpenMPI and OFED software interfaces. Supported products include Curtiss-Wright's CHAMP-FX3 Virtex6 FPGA board, the CHAMP-AV8 dual 2nd Generation Core i7-based multiproccessor board, the VPX6-1956 2nd Generation Core i7 SBC, the VPX6-6902 SRIO/Ethernet switch, and the VPX6-490 GPGPU module. OpenVPX enclosures are supported including a small 5-6 slot and 19" rack 16-slot air-cooled Chassis.