In the chip race, efficiency beats speed

Following is essentially the content of today’s Green Tech Newsletter. I generally don’t post the newsletter verbatim in my blog, but it seemed important enough in the context of my piece about Barcelona. The newsletter, by the way, is free. You can subscribe here.

AMD and Intel have been locked in fierce races for quite a while now, the latest being the race to first deliver a quad-core processor. Intel managed to win by a nose there, unleashing Tigerton a few days before AMD presented Barcelona.

But speed alone won’t be the determining factor as to which of these two processors will reign in the datacenter. Driven by customer demand for more energy-efficient hardware, both companies are talking up the importance of a less-familiar metric called performance per watt (PPW). As Intel tells it, “performance per system watt is calculated by taking the performance score from a benchmark or application and dividing it by the average system power usage (AC power from the wall).”

At a press meeting with Randy Allen, VP in charge of AMD’s server and workstation division, Allen confirmed that there’s been a shift in what customers want from hardware. “The traditional buying criterion has been peak performance. People would buy at the highest peak performance they could or they would buy on performance per dollar. … This emergence of performance per watt has been dramatic over the last two years.”

The problem with PPW is it’s difficult to measure. To draw on a familiar example, what’s a better performer: a hybrid or a pickup? The hybrid will certainly deliver more miles to the gallon if you’re using it to get around town for work or play, but if the task is hauling lumber or heavy equipment, you’ll get better results with the truck. Then again, if speed truly is what you need, gas costs be darned, maybe that Porsche is right for you.

Back to servers, then, how does a datacenter operator go about assessing whether Machine A or Machine B will deliver better PPW for, say, the company’s accounting application? In my interview with Bruce Shaw, director of server operations at AMD, he told me the best approach is to load up your app on your server, then measure it at the wall.

Sure, that’s not as simple as being able to look at the vendor-provided figures and know for certain which machine is the best overall performer for your specific needs. Then again, the exercise is a valuable one, given that filling your racks with the machines that gives you the best PPW will save you money in the long run — just like filling your corporate fleet with hybrid trucks just might make more sense economically and ecologically than a fleet of Segways.