Asymmetric Core Computing
March 30, 2010
The Army needs mobile, adaptable, cost-effective, high performance computing solutions to facilitate efficient processing for increasingly digital and networked battlespaces.
The binary computing field is rapidly evolving, driven largely by market forces and physical limitations in current processor fabrication techniques. This evolution has the potential to provide scalable processing power to tackle many of the Army's processing needs. However, the emerging processor options are asymmetric and various computing cores often have differences more pronounced than their similarities.
The Army Research Laboratory is engaged in novel algorithm design, analysis, and hardware control strategy research to provide unique solutions within a complex asymmetric core computing approach, thus bringing supercomputing levels of performance to Soldiers and commanders in operational settings. ARL does this by focusing on a methodology that uses resources such as task parallel, multi-core computer processing units (CPUs), customizable circuit devices such as Field Programmable Gate Arrays (FPGAs), and highly multi-threaded data parallel computing assets such as Graphics Processing Units (GPUs). By focusing on commodity resources, ARL's paradigm remains flexible and low-cost throughout the lifecycle.
A customized hardware approach is also being developed to field small footprint asymmetric core workstations with high floating-point operation performance (FLOP). Customized designs in the size of a standard workstations are leading toward FLOP performance approaching or even surpassing supercomputing clusters of von Neumann CPUs that are usually stored in large computing facilities.
ARL's research team is developing new approaches for algorithm formulation and optimization in the complex tasks of distributing and mapping computational requirements to an integrated runtime system, which features asymmetric core resources.
ARL succeeded in developing real-time signal and image processing approaches for obstacle avoidance and concealed target detection.
ARL's asymmetric core solution is executing 610 times faster than the baseline Matlab implementation, and 38 times faster than single-core C implementation. Electromagnetic wave propagation models have been developed for asymmetric resources (40 times over single-core CPU solutions) that will allow for commanders to better plan and optimize mobile ad hoc communication networks. Real-time ballistic threat assessment and reduction models are being developed by blending ray tracing approaches for asymmetric cores and urban/terrain models.
These demonstrations represent a significant breakthrough in advancing both the capability and the capacity of computing in operational environments. New applications and opportunities continue to emerge at every level.