With growing pressure on firms to reduce the carbon footprint of their datacentres, and vendors falling over themselves to brag about how green they are, it is a little disheartening to learn that the problem of processor power consumption is getting worse rather than better.

A recent study commissioned by AMD reveals that while overall server power consumption has remained flat for several years, the power consumption of cheaper volume servers actually went up. This type of system, which typically comprises the x86 processor architecture and either Windows or Linux, accounted for 96 percent of servers sold in the US in 2005.

What is more, all the signs suggest there will be a boom in demand for volume servers over the next few years, largely for reasons other than cost. Some AMD executives have suggested that a widespread enterprise migration to Linux and distributed environments could play a part in boosting server demand, for example.

Virtualisation will also have a big impact on power consumption because it requires servers with powerful processors and plenty of memory. The code bloat that afflicts many mature architectures is also fuelling demand for high-performance kit.

Either way, there will be more chips inside each box dissipating heat. Perhaps servers should be sunk in tubs of water to provide central heating or at least boil the water for tea and coffee. That way there would be some payback.

More seriously, when servers account for a global power bill of $7.2bn and rising, it is reasonable to assume there is a problem. So what can be done? First, some tenets of enterprise computing may need to be challenged.

For example, there is much talk of better “performance per watt” in processors – but what does “performance” actually mean? If it defines the effort a system expends to manage bloated applications in grossly overblown operating system environments that barely advance real business functionality from 10 years ago, then a rethink is needed.

So it could be time to start thinking the unthinkable, to put sentimentality aside and seriously consider retiring the x86 processor architecture for good.

Intel’s plans for tera-scale computing suggest it has already reached this conclusion. The initial target, already achieved in the labs, is to get an 80-core single-chip parallel-processor device to run at one teraflops on less power than many single-core x86 devices currently in service.

Where might this lead? Well, for a year or two, nowhere special, but it was only 11 years ago that the first one teraflops performance was achieved on the ASCI Red supercomputer running 10,000 Pentium Pro processors.

If something tangible comes out of tera-scale, it could seriously dislocate current IT strategy and infrastructure. But the temporary chaos that tera-scale would cause may by that point be a less painful prospect than any further extrapolation of x86.