For mainstream microprocessor manufacturers, this decade is what the 1970s was for automakers: a time of reckoning with energy.
For Detroit, an oil shortage drove consumers to choose fuel-efficient Japanese cars over American wheeled battleships. For Silicon Valley, today’s problem is the excessive power consumed by chips traditionally designed purely for speed. Consumers and business users want mobility. Notebook computer sales overtook desktop system sales for the first time in May 2003, according to The NPD Group, a research firm. More of the chip market is now delivering strong performance coupled with acceptable battery life in small packages. Meanwhile, IT managers are straining custom air conditioning systems to cool racks of scorching servers. They, too, want performance, but with reasonable power costs.
When Intel last year cancelled the Tejas chip, an even hotter, faster desktop processor than the current Prescott design, it acknowledged that no one can avoid taking power ramifications into account. Intel’s and AMD’s engineering attention has turned to innovations that will tone down their chips’ thirst for power. For laptop, desktop and server processors, the companies are now rolling out fundamental design and materials advances to hold down Watts and degrees while still delivering adequate performance.
Philosophical differences
The change has required re-educating engineers. Mooly Eden, a director in Intel’s mobile products group, describes in vivid terms the radical shift in engineering culture that led to the popular, low-power Pentium M and the Centrino brand of mobile and wireless chips. “If you’re a normal architect or engineer, the way to have your name last for eternity is to build a microprocessor with the best performance ever,” Eden says. “The big revolution in the design of the Centrino and the new microprocessors was: ‘Power is God. Give me the best performance you can within a power envelope.’” Engineers didn’t always take kindly to Eden’s rejection of ideas that boosted performance but also power consumption. Doing so seemed a huge insult to engineers steeped in a culture in which maximising performance was the path to glory.
Not that the idea of compromise was exactly unprecedented. Back in 2000, Transmeta emerged from Intel’s shadow in Santa Clara, Ca, preaching the gospel of low-power design for PC processors. Transmeta won attention and acclaim, but the company’s techniques inhibited peak performance, analysts say. Meanwhile, after Intel developed its own low-power chips, it seized the markets Transmeta had pioneered, says Gartner Group analyst Martin Reynolds. This January, amid continued losses, Transmeta announced that it would scrap making chips in favour of licensing its technology. But don’t misread the company’s troubles, Reynolds says. Low power remains in vogue. It’s just that performance never went out.
Wringing out the juice
What has the change in philosophy meant in practical terms? A lot. Engineers have tackled the power problem on the level of basic materials and individual transistors, on the level of circuits and on the level of entire systems. Their work has led to significant advances that have either increased performance without increasing power or reduced the ways systems waste power. Both Intel and AMD now produce laptop processors whose speed is well over a gigahertz but whose peak power consumption is around 30 W. In the late 1990s, a 30-W processor had a top speed of only 600 MHz.
To accomplish this, Intel, for example, has embarked on a sweeping campaign to plug “leakage” of electrical current that is responsible for chips wasting power and emitting heat even when they are nearly idle.
On another front, AMD teamed up with IBM in 2003 to build new transistors that can work faster than traditional ones at the same voltage. Their partnership came to fruition at the end of last year, when they announced that they would begin making processors using strained silicon and silicon-on-insulator materials technology. Silicon-on-insulator means building a transistor atop a sandwich of pure silicon, an insulator and more silicon. The sandwich lowers the capacitance of the transistor, meaning that it switches more quickly, resulting in faster operation and uses less power in doing so. Strained silicon involves stretching or compressing the spacing of silicon atoms along the path where electricity flows in a transistor. For different kinds of transistors, the stretching or squashing enhances the current flow in the silicon, again increasing operating speed without an increase in voltage. AMD expects to begin using the technologies in chips this year.
Chip architects at both AMD and Intel have progressively given processors more ability to fine-tune voltage and power, so that performance matches—rather than wastefully exceeds—what users need. For example, a typical Microsoft Word session requires only about 10 percent of the capacity of a processor, so it would be decadent for a processor to run at peak voltage then. AMD’s PowerNow technology, which debuted in chips for laptops in 2000, can automatically move among 32 different voltage levels to vary clock speed according to system demand. The company rolled out similar Cool ‘n’ Quiet technology for desktops in 2003 and this year will include PowerNow with server processors, says AMD divisional marketing manager Bahr Mahony.
Intel’s competing SpeedStep technology originally throttled voltage and clock speed between two levels, but the Enhanced SpeedStep, rolled out with the Pentium M in 2003, offers finer-grained control. Intel launched the server version, which it calls demand-based switching, in 2004. The Pentium M also includes a deeper take on the idea of power and performance on demand, Eden says. Traditionally, every functional unit in a processor would be powered until the power management system discovered which units were idle and turned them off. Now the default is for blocks on the chip to be off until they are needed, he says.
Transistor shuffle
Yet another way both Intel and AMD have attacked the power problem is by rethinking how best to spend the “transistor budget”—the hundreds of millions of transistors they can fit on a die. Moore’s Law predicts that the number of transistors on an integrated circuit will double every 18 months. Chip designers have traditionally used those extra transistors to build an even more complex, potent and power-hungry core than before. The result was a chip that ran faster and hotter. The next leap for Intel would have been from Prescott to Tejas, which was simply too hot.
Intel could make the decision to cancel Tejas, analysts say, because the company had finally developed a viable alternative: using its growing transistor budget to put two somewhat simpler cores on the same die. Sound like a sacrifice? It isn’t really. Yes, each core in a dual-core chip runs at a slower speed than a single-core chip, but it yields higher overall performance, because two weaker cores can lift a heavier load than one muscle-bound one, says director of circuit research Shekhar Borkar.
Another use of the transistor budget in the name of power is to integrate other kinds of circuits onto the same die to create a “system on a chip” (SoC). That’s nothing new, but the closer circuits are to each other, the less power they need in order to send signals back and forth. So although chip makers haven’t always had power in mind, they have been integrating more things onto the same die as processors for years, says semiconductor analyst Jim McGregor, of In-Stat. PC microprocessors used to have separate maths processing units and caches. Now those are integrated. AMD saves power by also integrating a memory controller on its chip, a move analysts say Intel should make as well.
Borkar believes that Intel should integrate all kinds of highly specialised, power-efficient cores onto its processor dies to save even more power in a computer. Calling the idea “integration to the extreme”, Borkar envisions routine functions, such as MPEG media encoding and decoding, TCP/IP network processing and even handwriting or speech recognition, in cores sitting alongside general-purpose logic. Such integration would boost performance while saving cost, space and power.
Over the years, chip makers have invested so much in performance that processors are now more than capable of doing the tasks consumers and business need them to do, says McGregor. The market’s new attention to power efficiency seems to reflect this satisfaction with performance, and that, in turn, has altered the philosophy and direction of Intel and AMD, which are now racing each other not just to raise speeds but also to lower wattage. By accounting for power consumption, they will enable their customers to do tasks the way they want to do them: coolly and on the road.