For years, the guts of a PC remained largely unchanged. In one product generation after another, Intel processors connected to a chipset that consisted of a memory controller and an I/O controller. That's about to change.
If Intel has its way, the functions carried out by these chips, and other components inside the PC, will soon end up inside the processor.
"You're going to have a hard time knowing where the processor ends and where the system-on-chip begins," said Pat Gelsinger, senior vice president and general manager of Intel's Digital Enterprise Group, during an interview at the Intel Developer Forum conference in Beijing last week.
In Intel's vision of the future, the processor becomes the system architecture instead of a component. "You're going to see incredible integration in the future," Gelsinger said.
The first hint of these changes came five years ago, when Intel added simultaneous multithreaded processing capabilities to boost performance in its Pentium 4 chips. Until that point, the basic architecture of most computers – a single processor core running a single thread of instructions – had remained unchanged for decades.
More change came with the introduction of Advanced Micro Devices' (AMD's) Athlon 64 processor, which integrated the memory controller with the processor on a single die. This first step towards integration of the processor with other PC components was followed in 2005 with AMD's introduction of dual-core processors, and the 2006 release of the first quad-core chip from Intel.
Building on these advances, Intel engineers are designing processors with more cores that integrate many of the functions assigned to other chips. As the number of cores increases, so does the number of threads that can be processed simultaneously, opening the door to further performance gains.
In 2008, the company will ship its first Nehalem processors, which will be manufactured using a 45-nanometre process and include an integrated memory controller, eliminating the need for a front-side bus.
At roughly the same time, Intel will begin shipments of Tolapai, a system-on-chip that includes an x86 processor, integrated chipset, and an encryption coprocessor for server appliances. A similar chip based on components for consumer electronics, such as integrated graphics, is also slated to enter production next year.
These chips are a hint of what Intel has in store for the future. "You can expect to see us do a wider range of system-on-chip solutions," Gelsinger said, noting that many applications, such as mobile computing, are well suited for these types of chips.
And more is coming. Intel's Larrabee chip, now under development, will pack dozens of x86-compatible processor cores onto a single chip. Billed as an alternative to general-purpose graphics processors, Larrabee will be a programmable and highly parallel processor capable of processing graphics and visualization workloads, such as fluid modelling and ray tracing, Gelsinger said.
In time, these functions could also be integrated into the system-on-chip processor, eliminating the need for external graphics chips.
The changes underway at Intel will have repercussions beyond the PC. The company wants to see system-on-chip processors based on its version of x86, called Intel Architecture (IA), find their way into every computer imaginable, from the smallest handheld devices to the largest supercomputers.
"We're taking IA to very small things and to very big things, in terms of capabilities and performance," Gelsinger said, adding that "all other architectures are marginalized."
While the shift towards system-on-chip processors will yield more powerful chips that are cheaper and consume less power, the chips will be less flexible. They cannot be mixed and matched with other components in the same way that today's processors can. As a result, chipmakers like Intel will have to produce a wider range of products, each designed for a specific application or type of device.
"If people think we have a large number of SKUs (stock keeping units) now, wait until we get into that era," said Justin Rattner, Intel's chief technology officer.
To produce such a wide range of products efficiently, future processors will have to be modular, allowing chipmakers to quickly "dial in" the number of processor cores and other components, such as memory subsystems and graphics, required for a specific chip, Rattner said. "All processor companies are going to have to figure this out," he said.
Intel is already taking steps in that direction. Nehalem is the first processor architecture Intel has designed with some degree of modularity in mind. "Nehalem is certainly not the ultimate in integration, but it does demonstrate that people are starting to design in a modular way," Rattner said.
"That's a big change," he said.