Multi-core provides user-friendly design opportunities

While chip design moves deeper into multi-core, I see two trends that should provide the user an altogether different experience in getting more from the chip than before. Hitherto the CPU design inhibited the user from getting the best out of CPU cost and power. Take the cost factor. The rapidity with which chip techology advanced to keep up with Moore’s Law deprived designers of significant opportunity to think in terms of investment protection. CPU makers rapidly phased out their CPUs to introduce an advanced one. But now both Intel and AMD are making multi-core chips that are “expandable and scalable.” This means that the microarchitecture can be enhanced from the 2-core to more cores, thus giving the current generation of chips almost the same power as a future chip with more cores once the present chip has been “expanded and scaled.” Intel’s “Caneland” chip series is specifically being designed for “investment protection.”

VIRTUALIZATION
The second trend is virtualization. The term virtualisation dates from 1960s. IBM M44/44X machine was called a virtual machine. Earlier, virtualization was a design concept contrasting with transparency. According to one definition, a virtual artifact is visible, perceivable, but does not physically exist, whereas a transparent one exists physically, but is invisible in use. Today, virtualization has an extended meaning of making a single physical resource appear to function as multiple physical resources, enabling the same person use that resource in multiple ways or many persons use that resource independently of one another. In this sense, virtualization is in its infancy. Virtualization in CPUs gives users the ability to use the CPU to much greater power than hitherto. Not many users are aware that with conventional CPU designs even though CPU forms a major chunk of the price they pay for a computer, they hardly use 10 percent of the CPU power.

Both Intel and AMD have added virtualization to their chip designs. Intel calls it Virtual Technology while AMD calls it Pacifica. Most of the design changes caused due to virtualization pertain to memory management. The two companies have enlarged caches that store data for virtual machines. This enables increasing the switching speed between a virtual machine and other devices on the server. A design idea that AMD is working on to get competitive advantage is moving memory management for virtual machines back onto the chip. With this design, the designer does not need the hypervisor—a kind of OS talking directly to HW—to manage the use of memory by each virtual machine.

Chip virtualization is forcing a major change in server designs. To host virtual machines, servers need a lot of memory, because each virtual machine gets an indivisible allotment. Server designers are still trying to figure out what exactly needs to be done to incorporate more virtualization-friendly features into their designs, but they know for sure that virtualization will enable more virtual machines per server, thereby giving the user greater cost-efficiency and power. Virtualization increases overhead by 15-20 percent because virtual machine software negotiates through an operating system to talk to the hardware. However, designs are emerging that enable the software talk directly to the hardware. This will bring down overheads, and designers are not far from achieving the goal of making negligible the difference between physical and virtual server operations.

MIGRATION
Future Intel motherboards will have a feature called “Flex” to assist in the migration of virtual machines across different servers. Enabling such a migration is a great design win. VMWare, SWsoft and HP offer migration but in a limited way. The ultimate design success in migration will come when migration occurs not only across boundaries between chips from the same manufacturers but also from different manufacturers. In other words, it should be possible to migrate from Intel- to AMD-based servers, and vice versa.

By 2010, Intel projects 25 percent servers will be running in virtualization mode, from the present around 5 percent.