Chips with everything
The rate at which processor technology is advancing is opening many new doors in the world of servers
Once again some big names in the IT world are talking about the notion of a computer on a single chip, aka “a system on a chip”. The idea has come and gone before, but has swung back into favour because chipmakers are able to squeeze increasingly huge numbers of transistors onto each chip. But the main problem is finding useful things to do with the extra capacity.
Intel chief technologist Pat Gelsinger talked at the recent Intel Developer Forum about how third-world countries and developing economies will need computers unlike those we currently use.
These new systems must be low-cost and low-maintenance, and many will work as routers or provide other network infrastructure for small rural communities. Offering them our old equipment simply won’t do. Nor will the specialist “system on a chip” such as Intel’s XScale offerings that are available now but are not x86 compatible.
Others are not necessarily convinced of this point, but Gelsinger argued that the poor quality of mains electricity in such countries would cause the power supplies in most of today’s computers to fail.
Either way, most people agree that a site visit for maintenance is an expensive option when the computer in question may well be located on top of a telegraph pole in the middle of the savanna. To operate on low power, such devices will need a small number of chips and state-of-the-art power supplies.
Un-coincidentally, the latest Linux kernel, released late last month, can support single-chip computers. It also has a feature to enable a new kernel to be loaded to replace an older one. Currently, you would need to reboot a Linux server to load a new kernel.
The most exciting development favouring single-chip systems was discussed by Gelsinger recently. He reckons that in five years we will have processors with 128MB of L2 cache on-chip.
Research also indicates that chipmakers will be able to attach another 400MB or so directly onto the CPU inside the chip packaging. This means the RAM would be able to work much faster than is possible using today’s architecture of RAM linked to CPUs via memory busses.
The result: a single chip containing a Xeon style processor and 512MB of RAM. That would require a phenomenal number of transistors on a chip, and so adding a few more to handle networking and other system features would proportionately not be a huge step.
Obviously such advances could be used to create a single-chip computer. They could also be used to make some pretty powerful servers for you and me. After all, AMD is selling dual-core server chips now, and Intel will start doing so later this year. Both plan quad-core versions in 2007.
By 2010 we could have eight or even 16 cores on a chip. That could mean a motherboard with 4GB or 8GB of RAM directly wired to the CPUs. I might be being short-sighted, but 4GB RAM sounds enough for most applications, and the potential performance of such a system would be stunning, provided that the software was designed for this kind of environment.