These doubts were exacerbated by the fact that modern applications – and the trend towards touchscreens, sophisticated user interfaces and larger displays with higher resolutions – place much greater demands on graphics, while keeping power consumption at less than 20W.
AMD is meeting these challenges with a fundamentally new approach – the Fusion architecture. This combines the cpu and gpu on one chip, along with an accelerated processing unit, or apu. The graphics processor is a general purpose gpu (gpgpu), which differs from a standard gpu through its flexible parallel processing units.
This means the gpgpu can also be used for general compute intensive, parallelisable operations, increasing non graphics performance. Typical uses include numerical mathematical applications and all types of encoding/decoding tasks and encryption and network packet processing in particular.
The Fusion architecture takes advantage of the fact that processes on a standard cpu are generally serial. Parallelisation can only be accomplished in multiprocessor systems or virtually through time splicing of individual, relatively large processes. The situation is different with a gpu: here, tasks are distributed over many small specialised engines, which are linked according to their respective tasks and which manage the various tasks in each time step in parallel. What is special about a gpgpu is that individual processor tasks are not hard wired; instead, tasks are freely configurable.
For software developers, this process remains transparent because the gpgpu engines are based on proven standards such as Microsoft's Direct Compute and OpenCL.
Since real task formulations are seldom purely serial, efficiency advantages appear when using a fused apu, which can process both serial and parallel tasks. Traditional cpu architecture and programming tools are usually less suitable for vectorial data models with parallel multithreads. However, the classical gpu is not the only application with obvious performance advantages due to parallel processing.
The AMD G-series platform with Fusion technology is the only one chip apu solution currently available. Experts believe that, for the demanding embedded sector with its complex applications, this technology will be as revolutionary as the introduction of multicore technology in the consumer market. The competitive advantages of the AMD Fusion architecture with regard to performance and energy efficiency are clearly identifiable. However, some rethinking is necessary.
Traditional cpu performance benchmarks, used to compare platform performance, are no longer meaningful. What will be decisive in the future is apu performance, which can differ in otherwise identical embedded systems according to whether the application has more or fewer parallel tasks or threads.
Standard 2d and 3d graphic benchmarks offer an indication of the apu's potential as they describe maximum performance in a single graphic mode. Figure 1 illustrates the performance of preproduction Fusion APUs in comparison to other processors for a special application (overall performance) and for a 3d benchmark. It should be noted that maximum power consumption can differ substantially.
ETX or XTX?
COMs are generally small embedded pcs with a standard form factor and additional core functionality. Because they can be mounted on customer specific baseboards, the development cycle can be reduced substantially by taking advantage of preintegrated core functions. Developers and manufacturers don't have to worry about special technologies; they can stick to specific interfaces and I/Os around standardised COM using familiar technology and their core competencies.
In this way, they can implement a new processor architecture, such as Fusion, through a simple module exchange with little additional effort.
ETX and XTX (for applications without an ISA bus) are implemented in millions of baseboards, so the question of which form factor to use for future COMs does not arise. However, for new projects and redesigns, it is also often wise to implement proven standards and the natural choice for new designs which do not require ISA is likely to be XTX.
For this purpose, congatec offers the ETX COM conga-EAF and the XTX COM conga-XAF. Both modules are based on the AMD G-series Fusion processors and the embedded controller hub Hudson E1 and are available with single core processors running at 1.2 or 1.5GHz, or with dual core processors running at 1, 1.4 or 1.6GHz and a thermal power dissipation of between 9 and 18W.
By default, the modules support four PCI bus masters (Rev2.3 compliant 33MHz), six USB 2.0 ports, two SATA II ports, two IDE channels (master/slave, up to ATA 133), a 10/100Ethernet interface and an HDA audio port with analogue support (line in/out and Mic) via ALC262. The modules have ACPI compliant power management and support Suspend to RAM (S3). Common embedded controller features such as i2c, watchdog, backlight and power dissipation are supported, as is the trusted platform module (Rev1.2) for safety critical applications.
The apu graphic, with two independent graphic controllers on the apu, supports vga (VESA compliant, up to 2560 x1600), lvds (single/dual channel 18/24bit), DisplayPort and HDMI//DVI (shared over FFC).
Good for embedded applications
The G-series processors are a good fit for the embedded market, especially for existing users of ETX or XTX modules. Used mainly in industrial and building automation, kiosk applications, medical technology and gaming as well as digital signage, these applications have in common the trend towards ever more realistic (3d) visualisation, with resolutions through to HD.
But today, high resolution touchscreen controls, regardless of whether they are used in machine operating terminals or kiosk systems, require higher graphic performance than could be offered by previous fanless solutions. Users can upgrade their designs by changing modules in order to enable more graphic intensive user interfaces. This leads to new possibilities in medical technology, particularly with regard to imaging technology, and analysis devices where the Fusion architecture can exploit the advantages of parallel processing via the apu.
In addition, the ratio of computing power to power consumption enables battery operated and/or fanless devices with high graphic requirements –often for the first time. ETX and XTX modules with Fusion technology will be available for many years, allowing easy entry into new levels of graphics performance and user friendliness.
The AMD G-series
Currently, G-series Fusion processors are available as 5+2 year long term embedded products with clock rates of 1 to 1.6GHz and with single and dual core apus, each with a 512kbyte cache per core.
Depending on clock rate and the number of cores, system target design power ranges between 9 and 18W. In addition, there is an ultra low power variant without a gpu. Apart from two independent display controllers, the apu provides four additional PCI-Express 2.0 lanes for system enhancements. The parts also feature 64bit wide ram connected directly to the apu.
The G-series platform is executed as a two chip solution. The AMD A55E1 controller hub, designed to meet the requirements of the embedded market, contains additional PCI, PCI Express, USB, SATA and HD audio interfaces. Both chips are connected simply and securely via four PCI Express 2.0 lanes.
Martin Danzer is development manager with congatec.
