Intel aims to lower IO power

ONE OF THE easier sounding projects that Intel showed off at its research day is scalable IO power research. As is usual with these things, easy sounding usually means hard to implement, and dropping IO power by an order of magnitude is just that.

Like all good research, it has a pretty graph, and this one is also pretty informative as well. What Intel boffins say is that bandwidth needs are increasing exponentially, and so is IO power. An order of magnitude better is equal to about five years worth of scaling.

That sounds good, but how? There is not one single answer, but a lot of the techniques mirror the power savings seen on modern CPUs, at least on a macro level. The idea is to ramp power up and down with speed, and as always the devil is in the details.

The way to do it is called Adaptive Power Tuning, an APT (Ha! Geddit?) acronym. The general idea is just like overclocking, the higher speeds need higher power, so when you need all that speed, you crank the voltage up and burn power to get the data out. When you don't need the speed, you lower voltage, lowering the effective link bandwidth.

The goal is to run the link at as close to 100% of the available bandwidth as you can. When you need more bandwidth, you simply crank up the voltage effectively raising the cap, and run it at 100% of the new cap. That way only the needed energy gets burned, they call this part Demand-based IO.

Intel can tune the supply voltage, that is the basis for a lot of this, as well as the transmit swing voltage, bias and several other parameters. Think Speedstep for IO links and you are in the right neighborhood.

The last bit is on chip Power-free Interconnect Compensation. The point here is to be able to compensate for distorted signals, light duty signal processing, with little or no added power. This way you can compensate for the hiccups caused by voltage and speed swings.

It is a closed loop system, the chips tune themselves, ramp up and down, and do their thing without OS or user intervention. They know how much work they need to do, and set parameters in such a way as to do only that.

In the end, you hopefully get that order of magnitude increase. As you can see, they are at 14% the power of PCIe with similar data rates right now, so it is in range. With luck, this research will buy us the aimed for five years. µ