This article was published in the September 2015 issue of Maximum PC. For more trusted reviews and feature stories, subscribe here.
Demystifying Processor Power Consumption
My recent column on AMD’s advanced voltage/frequency scaling confused some readers. It’s worth tackling this subject again, because power consumption is the most challenging problem for chip designers.
It’s also a challenge for users who struggle with thermal requirements when building custom systems. The root problem is that each new generation of chip-fabrication technology enables smaller transistors, so designers cram more of them on each chip. It’s great for processor performance, because we get more powerful CPU cores, more cores per chip, and more cache. That’s the bright side of Moore’s Law. The dark side is power consumption. Although today’s transistors are extremely power efficient, their wattage adds up when a chip has billions of them.
Most home users could afford to pay the slightly higher electric bills if their PCs consumed twice as much power. For data centers with thousands of rack servers, the higher bills would be more deadly. But heat dissipation—the corollary of power consumption—is the real killer.
Smaller transistors can lead to impressive designs, but power draw is important as well.
In the early 2000s, chip designers hit the “power wall.” No longer could they gain performance simply by cranking up the clock speed. Chips were getting so hot that they stopped working. So, designers found alternatives, such as spreading the software workload across multiple CPU cores. Other powersaving techniques are dynamic voltage/frequency scaling (DVFS) and advanced voltage/frequency scaling (AVFS). Both work on the principle of reducing the processor’s supply voltage and clock frequency when full speed isn’t needed. But the commonly used power equation I quoted (P=fV2) confused some folks. It says that power varies linearly with clock frequency and quadratically with voltage. For example, reducing the frequency by 50 percent also reduces power by half, but reducing the voltage by 50 percent reduces power by 75 percent. Voltage matters more than frequency because it’s a squared factor.
You can’t use this equation to calculate actual processor power, however. (For one thing, I deliberately omitted a capacitance factor to more clearly show the voltage/frequency relationship.) So, if your 1.2V processor runs at 3GHz, you can’t multiply 1.2 x 3.0 to find the wattage. The simplified equation merely expresses the relative voltage/frequency relationship.
Also, the frequency factor doesn’t apply to other things. One reader joked that his desk lamp would burn 864,000 watts because it uses 60-cycle alternating current at 120 volts. Luckily, AC frequency isn’t the same as processor clock frequency, or we’d still be burning candles.
DVFS is getting more complicated as designers divide their chips into more power, voltage, and frequency domains. Instead of varying the entire chip’s voltage and frequency, individual sections are controllable. Some domains can vary their voltage; some can vary their frequency; some can vary both; and some can shut off their power altogether. One company (Sonics) recently announced a power-management subsystem that can do all this stuff for dozens or even hundreds of domains. The trick is to vary these factors on the fly without noticeably hurting performance.
“Dark silicon” is a popular industry buzz term. It means that more circuits in your processor will power down when they’re unneeded, spring to life when duty calls, and then go “dark” again to save power. It’s like switching off the lights in your home’s unoccupied rooms, except it happens automatically, many times per second—and if it’s done right, you won’t notice a thing.
Tom Halfhill was formerly a senior editor for Byte magazine and is now an analyst for Microprocessor Report.
From maximumpc
from http://bit.ly/1Mg3kyw
