AMD offers special versions of its AMD64 microprocessors for embedded systems – that is, for using inside special-purpose devices, instead of general-purpose notebooks, desktops and servers. In the first part of this discussion, we examined the philosophy behind embedded devices, and the move away from proprietary processors to those based on the x86/x64 industry standards. We also looked at some of the advantages that special embedded versions of the AMD Opteron™ processors bring to embedded systems, and touched on sockets and longevity and pricing policies. In this second part, we'll dig deeper into the specifics of these processor offerings, and touch on power management and the Open Platform Management Architecture (OPMA).

Embedded software developers don't just write software. They target their software, including applications and infrastructure code, for specific hardware devices, or for classes of devices. That's quite a change from standard desktop/server development, where there's a lot of abstraction. That's why an entire ecosystem – actually, many huge ecosystems – have developed for the embedded world, with its own hardware and software makers, tooling, even conferences and magazines.

AMD's high-end x86/x64 processors have to fit into established embedded ecosystems, as well as gain support from existing embedded solutions providers. An important way it did that was to offer t the embedded versions of its processors – without any changes that would affect customers – for a minimum of seven years. Typically, a particular version of a standard desktop or server processor is around only for two or three years, and becomes obsolete when faster versions become readily available. While that's fine for the fast product cycles of general-purpose computers, that's not acceptable in the embedded world.

Fortunately, there wasn't much that AMD needed to create an ecosystem for the AMD Opteron embedded processors. On the software side, not much needed to be improved because AMD uses industry-standard instruction sets and operating systems (such as embedded Linux and Windows XP Embedded), tools like the GNU Compiler Connection, Eclipse and Microsoft Visual Studio can be used to create embedded applications.

On the hardware side, AMD has announced many partnerships. For example, even before AMD acquired ATI, ATI's RS4800 SVGA bridge and SB400 southbridge chips supported AMD64. AMD offers its own PCI-X bridge chip, but one is also available from Alliance Systems. PathScale offers an InfiniBand bridge for AMD64; Xilinx and Altera offer FPGA chips. There are also southbridge chips from NVIDIA, Broadcom and Via Technologies and more. The point is, the AMD embedded processors are widely supported for the market segments that they're addressing, such as single-board computers, blade systems, communications systems, and storage systems.

AMD's embedded Opteron processors are more than just "technology" – they employ real-world technology which is steadily being deployed in real-world systems. You'll find embedded AMD Opteron processors in systems like Agami Systems' new storage servers, Adaptec's SnapServer 510 network attached server, the Sun Netra CP3020 blade and numerous systems coming out of Huawei Technologies in China.

AMD Opteron processors are the flagships of the AMD high-end processor fleet. (There are also embedded versions of the AMD Sempron processor and AMD Turion mobile technology.) As with the desktop/server versions, they are available in three flavors: the 100 series, for single-processor boards; the 200 series, for dual-processor boards; and the 800-series, which support four and eight processors in a system. All are available in both single-core and dual-core versions, and it's reasonable to expect that there will be quad-core versions available soon – making for a very powerful embedded device. As with all AMD64 processors, you can boot it up with either 32-bit or 64-bit operating systems, and if you use a 64-bit OS, it can run both 32-bit and 64-bit apps simultaneously.

100 Series. I would expect that the AMD Opteron 100 series is popular in many embedded designs. These processors use AMD's 940-pin socket, and are available in a variety of clock speeds, with TDP (total power dissipated) from 30 watts to 104 watts. As of late November 2006, there are three single-core versions, at 1.8GHz, 2.2GHz and 2.6GHz. There are also two dual-core versions, both at 1.8GHz. (One of them is 55 watts, the other 95 watts.)

The Opteron 100 Series processors are as powerful as the desktop/server versions of these chips, with 1MB ECC L2 cache, and a 16-byte ECC integrated memory controller capable of working with DDR200, DDR266, DDR333 and DDR400 memory. There's a maximum of 32GB RAM, implemented in up to eight DIMM chips.

For high availability, you can configure it with hot-spare memory failover so the system will use redundant RAM if a memory chip fails.

The processor also includes three non-coherent HyperTransport Technology links, which can be used to hook up to a variety of I/O devices and buses. (Non-coherent doesn't mean that the links babble – it means that those links are designed for I/O. Coherent HyperTransport technology links can be used for either I/O or for processor-to-processor communications; the coherency means that the processors can maintain cache coherency over the link in a multi-processor configuration. A coherent link can also be used for I/O, instead of multiprocessing – system designers have a lot of flexibility.)

You'll find the 100 Series processors in dense rack-mounted devices, network appliances, storage systems or other locations where density is important – and where scalability is important as well.

200 Series. These processors are very similar to the 100 Series processors; the biggest difference is that they have one coherent and two non-coherent HyperTransport technology links. Therefore, you can use these processors in dual-processor configurations. Otherwise, they have the same memory capabilities, same 940-pin socket, and same thermal characteristics as the 100-series AMD Opteron processors. As of late November, there are three single-core versions, at 1.8GHz, 2.2GHz, and 2.6GHz. There are two dual-core versions, both at 1.8GHz, one of which runs at 55 watts, the other at 95 watts.

You would expect to find the 200 Series processors in higher-end applications, where there's more of a demand for threading. Dual-processor blade servers strike me as the sweet spot for these chips. They're also useful for devices that require a lot of I/O. For example, you could set up a dual-processor system to hook up multiple PCI Express buses. Each of the processors can dedicate a HyperTransport technology link to its own four-lane PCI Express controller, like the NVIDIA CK8-04 – giving you an aggregate of eight PCI Express lanes, each supporting up to 1GB of payload bandwidth.

800 Series. You guessed it: These chips are essentially identical to the 100 Series and 200 Series processors, except that all three HyperTransport links are coherent. That means that these processors can be used to "gluelessly" build up to eight-socket systems. Beyond that, you'll find the same speeds and feeds as on the other processors – and the same socket and thermal characteristics.

Where are you going to find these big guns? In large-scale devices, such as high-end network storage and other devices where you need a lot of processors – as many as 16 cores! – but you want to keep the chip count down in order to simplify the design, reduce costs, and operate with the lowest possible heat output. Remember, you don't need separate memory controllers, I/O hubs or "glue" chips for enabling multiprocessing; it's all built right into the AMD Opteron processors.

As of late November 2006, there were two single-core versions of the chip, one at 2.2GHz, the other at 2.6GHz. There are two dual-core versions as well, both at 1.8GHz, one running at 55 watts, the other at 95 watts.

Power management and the Open Platform Management Architecture (OPMA) are important features of the AMD Opteron embedded processors.

Many computer professionals think about power management as being something specific to mobile devices, like handheld computers or notebooks. That's one type of power management – dynamically tuning power consumption so as to preserve battery life. In the large-scale embedded domain, however, power management is a data center issue: when you pack dozens or hundreds of processors, hard drives, power supplies and other support circuitry into a data-center rack, it draws a lot of juice. What's worse, all of that power is emitted as heat – and you have to use more power to cool the systems. It's a vicious cycle.

Within the AMD Opteron embedded processors, there is a broad suite of features called AMD PowerNow! which manages the power. The plain old AMD PowerNow! is for mobile devices, but a version called AMD PowerNow! with Optimized Power Management (OPM) is what affects us here in the high-density data center.

OPM is AMD's technology for dynamically adjusting performance (and therefore, power utilization) based on CPU utilization. According to AMD, this technology can reduce CPU power during idle time by as much as 75 percent. We'll get into AMD PowerNow! with OPM in a future article.

The other important technology, Open Platform Management Architecture (OPMA – no relation to OPM, by the way), is an open specification for defining a common hardware interface between the embedded computer and its management subsystem. By using OPMA, along with the other open standards that AMD supports, application developers, hardware makers and software companies can leverage reusable components, simplifying the process of adding manageability features to embedded devices at low cost. We'll discuss OPMA in the future as well.

Alan Zeichick is a technology consultant and analyst who focuses on software development and microprocessor technology. Reach him at zeichick@camdenassociates.com.