Four core years... and then?

AMD's Barcelona keeps the company in the multicore race. That may not be the right game to play
Written by Leader , Contributor

Intel and AMD have been competing for 25 years. AMD was originally signed up by Intel to second-source the 8086 in 1982, and since then the battling companies have been in and out of more courts than Sampras and Agassi. Not that this matters to the market, which cares only about the best product at the lowest cost.

The latest battleground is quad-core, where Intel has stitched together two dual-core chips and got in ahead of AMD's "native" four-core Barcelona — which was announced on Monday. Both companies are justly proud of the achievement of putting four top-notch processors in one package, and the raw technological achievement is indeed worth celebrating — but it's futile unless it matches the needs of the market.

Like a powerful amplifier, unless the output impedance is the same as the load, that power won't be delivered. By concentrating on benchmark results, AMD and Intel risk losing sight of that. To some extent, benchmarks have always been synthetic — a guide to, not a guarantee of, performance in the real world. Now, the gap between the measured and the useful is increasing. More cores mean more capability, for sure, but for a decreasing set of applications. That's a very important set, especially for enterprises, but it's by no means the whole picture.

Moreover, while the marketing departments of chipmakers and OEMs will be pushing whatever measurements look the best, savvy IT departments making significant investments will know to ignore those in favour of their own tests. And while performance is of course important, so are many other areas where chip designers have been less inventive: security, usability, reliability — all have had far fewer transistors spent on them than their importance demands.

These three areas also help desktop and mobile applications, neither of which benefit that much from dual-core — far less from quad- or many-core systems. Neglecting such applications, or assuming that they'll do well enough purely through the natural reduction in power consumption and size that comes from scaling cores down, is not good enough.

All this may change, if the intensive research work on general-purpose parallel programming pays off. With no real sign of that happening soon, the intense focus on core multiplication "just because we can" risks leaving whole market segments stagnant. This risk increases with every generation that concentrates on doubling core numbers over architectural innovation. Both AMD and Intel know that new moves are needed. A move away from benchmark bingo would be a good start.

Editorial standards