March 12, 2007—UC Berkeley— In an invited talk at the International Symposium on Asynchronous Circuits and Systems, James Kajiya from Microsoft Research talked about "Signaling with conserved Quantities: Two Realizations in CMOS and Superconducting Flux Quantum Logic".

Kajiya is advocating a change from the practices of signaling with current flow to signaling with conserved quantities to address the MOS heat issue. Power consumption is becoming everyone's "hot" issue, even though the problem was identified in the early '80's. Everyone knows that mobile products need to operate at ambient temperature, since the total power budget cannot afford any high volume cooling devices.

In fixed systems, the issue is the limited cooling capacity available. Microsoft, Google, Yahoo and other large server farm operators are looking for sites that are at the junction of high bandwidth links and low-cost electricity to address the need for power to operate and cool the systems. The increasing heat density of the high-power components is almost matched by advances in cooling components. A look at in heat sinks over time show the changes from passive radiators, to forced-air coolers, to heat pipe and radiator assemblies. Even the most complex cooling components for microprocessors still cost less than $30, following the Moore's law curve of greater capacity at lower costs.

But even these high-technology coolers have physical limits. Designers must address the source of the heat, the circuits in the system. Designs would improve their dynamic power consumption efficiency by changing from a current flow to a charge flow for signaling. Power is consumed in the next-stage loads, interconnect lines, and active devices. The changing relationships of thresholds and supply voltages is making the challenges greater.

Static power is responsible for a major change in the industry. Intel cancelled a design when the leakage power for one chip exceeded their worst case assumptions. Because the only viable alternative was to use multiple cores running at lower individual speeds, this change signaled the end of the single thread microprocessor. Now, dynamic shoot-through and leakage currents are the dominating current consumers in ICs.

The industry has scaled supplies to keep the electrical fields in the oxide manageable, but now the transistors don't turn off. Traditionally, processes are designed for speed, by choosing the lowest possible threshold by setting the Fermi level about 2 kT below the threshold. The Fermi level is a function of the semiconductor doping levels. Sub-threshold leakage is a function of ∆Vg/Vt and these quantities are functions of kT/q, about 26 mV at 300°K. Unfortunately, the supply and threshold have converged.

One attempt to fix the problem is to change the gate oxide to a high k material. This works, except that you give up the main reason for choosing silicon as the base material in the first place. Silicon dioxide is easy to process and control. A high k material like HfO2 is coupled with low k interlayer dielectric materials and results in serious materials processing and structure issues.

Another set of fixes is to change from planar to 3-D FETs or multiple gate devices. These structures not only create implementation issues, but also will require a metallurgy that adheres to a vertical surface.

Here are a couple of alternatives: First, change the signaling to adiabatic logic to enable conserved quantity switching and lead to the ideal case of zero dynamic power. If circuits are designed to close the switch when the voltage differential is zero, or open the switch when the current change is zero, the circuit has no switching transients. Asynchronous adiabatic logic would use an AC supply that is half-wave rectified into two phases to supply the circuit power.

Another possibility is to consider scaling temperature. This alternative will cause designs for fixed and mobile devices to diverge even more. No one would be expected to carry a Dewar of liquid nitrogen to cool a portable device. The challenges for this alternative include the need for new transistors that are designed to operate at very low temperatures, and finding low cost sources of cooling to about liquid nitrogen temperatures. Although eventually this direction in ICs would take us to cryogenic systems, the first steps would be to higher temperatures. The advantages of operating at low temperatures include faster transistors, lower resistance in the interconnect lines, and much lower leakage.

Once you have made the transition to cold electronics, the next step is to change from CMOS to some form of superconducting logic, possibly based on Josephson junctions and SQUID (superconducting quantum interference device) type devices. One nice feature of these cryogenic devices is a logic inversion is accomplished with a transformer, and not an active device. Currently researchers are working on a microprocessor with a 30 GHz clock and circuits operating on mV-level signals at 100 GHz. The major limitation is the low level of integration possible in today's technologies.

To comment on this article send email to:gmoretti@gabeoneda.com