GPU solution hits 750GFLOPS peak per card slot

GE Intelligent Platforms has extended its rugged GPGPU (general purpose graphics processing unit) solutions family with the 6U OpenVPX NPN240 multi-processor. The NPN240 features two Nvidia CUDA capable GT240 96-core GPUs, enabling it to deliver up to 750GFLOPS peak per card slot, depending on the application. Multiple NPN240s can be linked to single or multiple hosts to create multi-node CUDA GPU clusters capable of thousands of GFLOPS.

The NPN240 provides a computationally-dense platform designed for demanding applications that are constrained in terms of size, weight and power. The relative ease of CUDA application development, together with the rapidly-growing infrastructure of software development environments and CUDA-knowledgeable programmers, promises to deliver considerable gains in productivity.

The rugged NPN240 is designed to deliver enhanced levels of performance in applications using parallel processing such as radar, sonar, and signal and sensor processing as well as image, video and graphics. Typical applications include stream processing of high-resolution images for deployed airborne surveillance. It is designed to be used in conjunction with the GE Intelligent Platforms IPN250 or SBC620, or any other OpenVPX-compatible host SBC (single-board computer). The IPN250 combines a single GT240 96-core CUDA GPU with an Intel Core2 Duo host processor operating at 2.26GHz and 8Gbyte of DDR3 SDRAM to deliver up to 390GFLOPS of performance per card slot.

The OpenVPX- and VITA48/REDI-compliant NPN240 is interoperable with other OpenVPX-compatible COTS boards to ease integration and reduce development time. Each CUDA-capable GPU node includes local DDR3 SDRAM as well as a 16-lane PCIe gen2 interface to the system backplane, providing maximum data throughput direct to GPU memory. Air-, spray- and conduction cooling build options are possible to allow deployment in the harshest military and aerospace environments.

The NPN240 will be available with GE Intelligent Platforms' AXISLib high-performance VSIPL DSP and math libraries for maximum code portability, enabling programmers to target their algorithms for the CPU or the GPU to optimize system loading and performance. In addition to Nvidia's CUDA environment, supported software includes Linux, Windows, Microsoft DirectX, OpenCL, OpenGL and MATLAB, together with Nvidia's PureVideo Technology and PhysX.

- Toni McConnel
Embedded.com