BY Simon Collins
Courtney Howard’s article in the July 2011 print issue of Military & Aerospace Electronics entitled “Harnessing high-performance computing” was further proof that, for developers of embedded systems for military and aerospace applications, high-performance computing (HPC) is a hot topic. That’s no surprise: Military and aerospace applications have long been voracious devourers of as much computing capability as they can find.
HPC in the commercial world is already well established, and its purveyors are learning to extract maximum advantage from many-core and multicore processors, as well as multiprocessor blades in application environments where a high degree of potential parallelism exists. The very highest performance computers in the world use clusters of thousands of the latest multicore CPUs (central processing units) from Intel along with thousands of Nvidia’s many-core GPUs (graphics processing units), often tightly coupled via InfiniBand and 10 Gigabit Ethernet remote DMA-enabled (RDMA) switched fabric networks with robust, high-performance driver support under Linux and Windows operating systems.
Commercial interest and appli- cation is also driving the development of an extensive infrastructure and ecosystem of supporting hardware, software, and middleware, reducing the need to develop complex, expensive proprietary solutions and further enhancing the attraction of HPC.
For military and aerospace developers, it’s all good news. Not only is the massive investment in HPC in the commercial world driving technology and performance developments at a spectacular rate (an example is the work being done by Nvidia and Mellanox on “GPU Direct,” a development that will greatly improve GPGPU, or general-purpose GPU, performance), but also it is doing so using the COTS (commercial off-the-shelf) principles that have become fundamental to embedded computing in military applications. The military and aerospace world is taking advantage of those developments.
Attracting substantial attention is GPGPU technology—applying the inherent massively parallel architecture of graphics processors to general-purpose computing tasks. GE was among the first to bring commercial, rugged GPGPU-enabled solutions to the market with the announcement in November 2009 of the GRA111 high-performance graphics board, following the signing of a unique, strategic agreement with Nvidia, the market leader in GPU technology, two months previous. The agreement gives GE not only direct access to silicon, but also the knowledge necessary to implement it appropriately for military and aerospace applications, as well as insight into Nvidia’s roadmap.
While commercial applications and military and aerospace applications have the same hunger for processing performance, their requirements are divergent when it comes to deployment. There is, of course, a world of difference between an air-conditioned data center and a military land vehicle when it comes to power consumption, heat dissipation, resistance to shock, vibration, and extremes of temperature, and so on. And when it comes to GPU technology, there is a similar world of difference between what a GT240 can be subjected to in a laptop and what it can be subjected to in an unmanned aerial vehicle.
In a laptop, the GPU will typically be configured via an MXM (Mobile PCI Express Module). This is akin to what a military and aerospace systems designer would probably think of as a PMC or XMC: a discrete printed circuit board—containing some specific functionality—that is mounted on a host board, rather than the functionality being integrated directly on the host board itself. In the case of laptops, the rationale for this approach is simple to understand: it allows manufacturers to implement alternative GPUs, or new generations of a GPU, without having to redesign the host board.
That’s not an approach, though, that lends itself to the rigors of military computing. It may work well in benign environments—and, in fact, later this year, GE will announce GPU-enabled products aimed at benign environment deployment and using the MXM architecture—but it is somewhat unsuitable for rugged systems given that it is designed for less challenging applications. It is in this area that GE’s relationship with Nvidia has proved especially beneficial: Access to in-depth Nvidia expertise has allowed GPU silicon to be implemented on, and integrated within, fully rugged computing platforms designed for deployment in the most demanding environments, conforming to the design, procurement, manufacturing, and qualification process standards required by military and aerospace customers, such as AS9100, IPC 610 Class 3, MIL-STD-810, and so on.
But “rugged” is not just about resistance to shock, vibration, extremes of temperature, and so on. A key area in which military and aerospace customers differ from their commercial counterparts is that performance per watt, not absolute performance, matters most—because, in an often physically-constrained environment, heat dissipation is difficult to achieve, yet absolutely vital if reliability in a mission-critical environment is to be guaranteed.
The maximum performance of the system is governed by the ability to remove heat from the processing units. That set of constraints has given rise to the notion of a solution’s SWaP (size, weight, and power) characteristics—its size, weight and power. Designing the most power-hungry silicon down onto the host board provides the best possible opportunity to design heat management systems to remove the heat most effectively, and keep the processors crunching numbers at maximum performance, thus optimizing SWaP for any given sub-system.
GPGPU technology is already being evaluated and fielded by a large number of military and aerospace programs—and with, for example, one radar application showing a performance increase of 15x compared with more traditional approaches, that’s not surprising. However, in an emerging market, it is important to realize that not all GPGPU platforms targeted at the military and aerospace market are created equal—and that not all claims to be “first” are 100 percent accurate.
In the excitement about the technology, it’s also important that program managers apply the same kind of selection criteria as they would for any other prospective solution, such as determining the nature of the extended product roadmap and the availability of long-term programs to support multi-year—perhaps multi-decade—deployments. As GPGPU technology enters the military and aerospace mainstream—as it surely will—those considerations will, inevitably, become second nature.
Simon Collins is product manager at GE Intelligent Platforms based in Charlottesville, Va.