Getting a view on medical imaging

In addition to powerful image display, extremely fast image capture and image processing functions are key requirements for embedded graphic applications. Sensor data have to be processed to generate image data that can be displayed at very low power consumption and, ideally, in real-time. This requires the highest possible parallel processing power. While existing solutions based on DSPs or FPGAs are relatively powerful, they are proprietary and new developments incur great cost and effort.

The ideal solution for graphics-oriented applications therefore combines all requirements in one compact and energy-efficient system: High multifunctional computing power; high parallel processing power for data processing and imaging in real time; high graphics performance for visualisation; plus platform and hardware independence for high reusability.

That's exactly what a heterogeneous system architecture with Computer-on-Modules (COMs) based on the new Accelerated Processing Units (APUs) of the AMD R-Series offers. They integrate an efficient multi-core x86 CPU for classic PC tasks and scalable workloads with a programmable vector unit for parallel computing tasks and high-performance computer graphics on one silicon die. Thanks to tight integration and specialiszed processing units, the APUs are very compact and extremely energy efficient.


Figure 1: The die structure of the AMD Embedded R-Series APU. Combining shared and dedicated resources ensures high performance with low power consumption

But which computing unit is responsible for the parallel tasks? The answer is simple; it's the integrated graphics processor. Fuelled by developments in the consumer sector — particularly computer gaming — graphics units have evolved to freely programmable specialists in parallel processing over the years. Modern GPUs now consist of several hundreds of processing units capable of performing complex calculations in parallel. They can do this with synthetically generated data from a computer game, but also with real data supplied by a wide range of sensors.

High-speed parallel processing

The integrated AMD Radeon graphics unit of the 7000 family is extremely powerful, offering betwee 128 and 384 graphics cores with a clock speed of up to 686MHz. Achieving a 3Dmark Vantage E result of 13,066, the AMD R-464L APU far exceeds the performance class of previous integrated graphics units on the market. For parallel computing tasks, the AMD R-464L APU achieves a maximum of 576GFLOPS single precision performance. COMs which integrate these APUs are therefore an ideal platform for demanding medical imaging applications.

To enable developers to make the most of this parallel processing power, the conga-TFS COM with AMD Embedded R-Series APU supports the latest, cross-platform APIs such as OpenCL.


Figure 2: The innovative architecture of the AMD R-Series APU integrates all major system elements — including x86 cores, GPU vector (SIMD) engines and unified video decoder — in a space-saving two-chip solution

OpenCL is a powerful programming environment that allows computing tasks to be distributed and processed across hardware within heterogeneous processor systems. OpenCL is special in that multiple parallel execution is possible at each step (SIMD = Single Instruction Multiple Data), which means that classic parallel computing architectures are also supported. This is crucial because besides graphics display many analytical problems are also ideal candidates for parallelisation.

Using parallel computing, many highly complex and accurate calculations can be performed in just a few computational steps where classic, serial CPUs would require up to several thousand steps. This obviously reduces computation times and energy consumption for complex computational tasks drastically. Medical imaging technology, with its pronounced, often well parallelisable analytics, can draw particularly great benefit from this increased efficiency. An example from an image registration application where stable video images are key illustrates just how great the efficiency gains from using OpenCL in heterogeneous system architectures can be: The OpenCL based algorithm executes 120 to 130 times faster than a classical calculation on a x86 CPU.

COMs enable developers and OEMs to design these new features particularly efficiently into their medical devices and applications. COMs integrate the core computing functions of a system on a swappable module as pre-integrated components, while external interfaces and peripherals are implemented on an application-specific carrier board that is relatively simple to develop. The separation of carrier board and computing unit is a distinct advantage when developing medical devices as these must meet numerous specifications, such as EN6061 which specifies an extremely low leakage current through external I/Os. This requirement calls for specific I/O expertise that is easily met in the design of the carrier board without having to adapt the more complex computing unit.


Figure 3: The conga-TFS COM with AMD Embedded R-Series

The COM Express module conga-TFS currently supports three versions of AMD Embedded R-Series APU ranging from the dual-core AMD R-272F APU to the quad-core AMD R-464L APU. The conga-TFS uses the AMD A70M Controller Hub and provides a powerful, compact two-chip solution with support for up to 16GB of dual-channel 1600MHz DDR3 memory.

The integrated graphics core supports DirectX 11 and OpenGL 4.2 for fast 2D and 3D imaging. A third generation hardware Universal Video Decoder provides seamless processing of H.264, VC-1, MPEG4 Part 2 and MPEG2 video streams. The choice of available graphics interfaces includes VGA and 18/24bit single/dual channel LVDS; three DisplayPort 1.2, one HDMI 1.4 and two single-link DVI for direct control of three independent displays are also available. Seven PCI Express 2.0 x1 lanes, one PCI Express 2.0 x8 link, four SuperSpeed USB 3.0 ports, four USB 2.0 ports, four SATA 6 Gb/s ports, a Gigabit Ethernet interface and high definition audio complete the comprehensive feature set.