32-bit processors have served us well since the 1980s, but their life is coming to an end. Once the 4GB memory limit (usually 2GB per process) seemed far away, but nowadays even home users could run into it.

The Athlon 64 (The Processor Formerly Known As ClawHammer) increases that address space to over 16 Exabytes (16 billion Gigabytes) of RAM. Which should be enough for the moment. Trouble is, you need 64-bit software to take advantage of it, and right now there’s virtually none. So it’s just as well that the CPU runs 32-bit software even better than the Athlon XP, thanks to new architectural features. High on the list is a DDR SDRAM controller (single channel, sadly), integrated into the core. This reduces the amount of data that needs to be sent over the processor bus, boosting performance and reducing latencies. And it’s one less component for chipset manufacturers to be concerned about, simplifying their designs.

Connection to the motherboard chipset is handled by AMDs HyperTransport bus, which also connects CPUs together in a multi-processor system – although this feature won’t be available for the desktop model. It’s scalable by adding extra HyperTransport controllers, but even a single ‘16-bits’ connection provides a bandwidth of 3.2GB/s in both directions (more than adequate, especially now the memory controller is integrated).

The move to a full 1MB Level 2 cache is another plus point. Other tweaks include adding two steps to the Integer pipeline, larger translation lookaside buffers (TLBs) and an improved branch prediction unit. There’s even support for Intel’s SSE2 instruction set, perhaps better known as NetBurst. Taken together they should provide a decent speed increase over the Athlon XP, and in time could provide Intel with some real competition.

Although the Athlon 64 improvements are welcome, these are offset by relatively low clock speeds. One reason for this is AMD’s use of a new SOI (Silicon-On-Insulator) fabrication process, but it could still prove a worthwhile choice — here’s why. As a processor works, its transistors are turned on and off regularly. But before a transistor can conduct electricity, it has to charge all its internal capacitance (its ability to hold electrical charge). This takes a long time in CPU terms, so reducing that time period could greatly improve performance.

SOI involves building the transistors on a thin layer of silicon, which in turn sits over an insulator such as glass. This cuts down the transistor capacitance, cutting the ‘turn on’ time, and making your processor faster. Just how big is the speed improvement? It’s impossible to say for the Athlon 64, as there are no non-SOI equivalents to measure it against. But IBM carried out some tests which suggested a 25 to 35% speed boost was practical. And with SOI also potentially offering to cut CPU power requirements in half, or even more, it’s a very interesting technology indeed.

Unfortunately it seems SOI is also a very tricky technology to get right, and AMD have been working on it for some time. The company had planned to use it with Barton core Athlon XPs, but problems forced them to change tack, and similar issues helped push back the Athlon 64 release date.

Still, it looks like AMD have finally got the process under control. If that’s true, expect the comparatively low clock speeds to be ramped up even quicker than anticipated, and the processor wars could take on a whole new dimension.