Heat Dissipation and the RSFQ Technology

The generation of heat in silicon chips is a problem at present, and is a problem that will get more difficult the more devices there are on a chip [1]. Large installations, as well as the uncomfortably warm contemporary laptop computers, already exhibit this problem.

There is a complete, but problematic, potential solution to this heating problem, suitable for large installations where expensive air conditioning is already required. The potential solution is the Rapid Single Flux Quantum (RSFQ) superconducting technology. Although the full details of this technology, purely based on nanophysical phenomena of superconductivity, are beyond the scope of this book, it is a wonderful example of the notion that the new laws of nanophysics may offer concepts for different, and possibly superior, devices.

An example, the performance of an RSFQ counting circuit [8] is shown in Figure 7.4.

Figure 7.4 Frequency divider [8] operates at 750 GHz, using RSFQ superconducting junctions. This device produces half-frequency output. Twice the output frequency, 2/out is shown superimposed onfin, both quantities referenced to left-hand scale. Bias current (abscissa) in mA does not signify heat, because the resistance is zero [8].

Figure 7.4 Frequency divider [8] operates at 750 GHz, using RSFQ superconducting junctions. This device produces half-frequency output. Twice the output frequency, 2/out is shown superimposed onfin, both quantities referenced to left-hand scale. Bias current (abscissa) in mA does not signify heat, because the resistance is zero [8].

The problem, for the RSFQ technology, is that refrigeration is needed to maintain RSFQ devices at the temperature of liquid helium. Liquid helium is a coolant widely used in Magnetic Resonance Imaging (MRI) medical installations. In addition to saving energy, the RSFQ technology allows extremely fast clock speeds, as seen in Figure 7.4.

The heat dissipation that occurs during the RSFQ switching events is described [8,9] as on the order of 10-18 J/bit, leading to an overall power dissipation estimated as 10-5 that of an equivalent silicon device. These devices are fast, as can be seen by the graph. These RSFQ technology devices have been considered for analog-to-digital converters, e.g., in missiles where fast data acquisition is urgently desired and refrigeration is already in place to cool infrared detectors.

The RSFQ technology is relatively simple, tolerating larger linewidths, but suffers from a primitive (underfunded) infrastructure for fabrication and testing. The active devices are niobium-aluminum, niobium (tri-layer) [10,11] Josephson tunneling junctions, which must operate well below the superconducting transition temperature of niobium, 9.2 K. Compact low power refrigerators that approach this temperature are becoming available.

The energy consumption and the cooling power (note that these quantities are additive, in terms of operating costs) needed for a large computer in this technology are less, by large factors, than for an equivalent silicon (CMOS) machine. The savings in power and floor-space are significant. It can be argued that the size of the overall machine would be much smaller because the design would not have to have the fins, open channels, fans, and air conditioners that are normal in large silicon computer/server-farm installations, simply to keep the devices from dangerously overheating, and to carry away the heat.

A benchmark in large-scale computing is the Petaflop computer, which carries out 1015 floating point operations per second. A typical modern Pentium-chip laptop computer operating at 1.3 GHz has a power supply rated at 54 Watts. A rough conversion of 1.3 GHz clock speed to 1 GFlop, would imply that a million laptop computers, if suitably coupled, might constitute a parallel architecture Petaflop computer. (This practical approach to making supercomputers has been implemented, as mentioned in the press, including use of Pentium computers and also SONY Plays-tationll devices).

On this basis, a Petaflop computer would consume 54 MW of power, or 1.7 x 1015 J/year. At a conventional rate of 10 cents per KWh, the annual power operating cost would be $47 M, not to mention the cost of the air-conditioning.

Binyuk and Likharev [9] offer the following projected comparison of the specifications for a Petaflop computer (1015 operations/s) in RSFQ technology, operating at 5 K, vs. silicon CMOS technology operating somewhat above room temperature:

In CMOS, one might have [9] 104- 105 advanced CMOS silicon chips, total power about 10 MW. Neglecting the air-conditioning needs, the operating cost would be around $8.7 M/y. The linear size of the computer might be 30 m, leading to a relativity time delay L/c = 300 ns. Such a large signal-propagation time is a drawback for a fast computer. This is a huge installation, on the scale of those in existence at major stock exchanges.

In RSFQ, one might have [9] 500 RSFQ logic chips and 2000 fast superconductor memory chips, at an estimated power at 5 K of about 1000 W. The linear size of the computer CPU might be 1 m, corresponding to a speed-of-light time delay of 20 ns. To achieve the reliable cooling of this 1 m "core" to about 5 K, a closed-cycle (recycling) liquid helium refrigerator would be required, at a total power of about 300 KW. Again neglecting air-conditioning, the annual cost is estimated as $263 000 per year.

This is a noticeably lower figure than $8.7 M/y estimated for the CMOS facility, to which one should add much larger costs in floor space and air-conditioning.

The ever-increasing energy cost of the pervasive silicon information technology was hinted at in the media discussion of the US power blackout of August 14, 2003. There were statements to the effect that the US needs an (electric) power grid that "matches the needs of information technology". "Server farms", typically housed in urban buildings with windows blocked in, that have appeared in information technology, are users of vast amounts of electrical power.

This costly power shows up as heat in the environment, cost on the balance sheet, and even represents a potential security threat, in its dependence on large amounts of (demonstrably insecure) electrical power.

1 THz

100 GHz

1 THz

100 GHz

petaflops computing, PeT computer i.8 nm 03 M"1 lMJJs 10MJJs

1.75 (im lOOKJJs

10 GHz

1 GHz historic trend

100 MHz i.8 nm 03 M"1 lMJJs 10MJJs

Si CMOS

1.75 (im lOOKJJs petaflops computing, PeT computer

1 GHz

.09 |im photolithography

0.065 urn 0045 Min

.09 |im photolithography

"no known solutions" (e-beam lithothography ?)

1996 1999 2002 2005

2008

2011

Figure 7.5 Estimated clock rates in RSFQ (upper) and conventional Si CMOS (lower) projected to the year 2011 [after 9].

Figure 7.5, adapted from Binyuk et al. [9], shows comparative projections of clock rate in RSFQ and CMOS computer technology. In the upper portion, "Nb RSFQ" refers to technology based on Nb-Al-Nb trilayer [10,11] Josephson junctions (JJ's), which operate below 9.2 K. In the upper right, "petaflops computation" and "PeT Computer" ("Personal Teraflop Computer", see [9]), with 10 million JJ's on a single chip, are predicated on an assumed 0.3 mm Nb technology. Note that 0.3 mm exceeds today's silicon technology linewidths, and is likely achievable with funding. In the far upper right, "HTS RSFQ", is a rough guess of RSFQ performance if the high-temperature superconductors could be brought into play.

The HTS superconductors are notoriously difficult to incorporate into electronic device technology, except in the simplest single-layer applications, such as high-Q radio-frequency filters for cellular telephone transmitting stations. For HTS superconductors the required refrigeration to 77 K is less costly.

In the lower portions of Figure 7.5 relating to the silicon technology [1], note that the linewidths by 2003, have already come down to the limit of photolithography. The very uncertain projections labeled "no known solutions" reflect that no mass production technology has appeared to replace the photolithography. These projections are based on the Silicon Technology Roadmap [1].

Other large issues relevant to Figure 7.5 are the greatly reduced power consumption in the RSFQ technology, coupled with its extremely inconvenient cooling requirements. These issues have been discussed in the text above.

It is clear that RSFQ technology will never find its way into wristwatches, laptop computers and most other basic computing venues for the silicon chip. The RSFQ technology may find commercial application in high-end installations requiring massive rapid computation.

The RSFQ technology offers a chance for much reduced power consumption and better performance in computing technology, at the major cost of learning how to make and use a variety of good refrigerators and making a large adjustment in mental orientation from silicon to superconductors like niobium.

To challenge and surpass a pervasive and successful technology, such as the silicon technology, is seldom achieved, and certainly not quickly. Jet airliners now outnumber propeller driven planes, and digital cameras may be displacing conventional silver halide photography. On the other hand, in spite of the complete success of nuclear power in the US Navy for submarines and aircraft carriers, private enterprise in the United States regulatory environment seems unable to make nuclear power work.

The RSFQ approach [9] offers an opportunity for innovators and entrepreneurs who are conversant with the advanced, extensive, and almost wholly untapped nano-physical science of superconductivity. To launch this radical computer technology, which is certainly technically sound; requires a largely new workforce and a start-up approach aimed at a niche market. There is a small core of dedicated workers competent in superconducting technology, including RSFQ, in laboratories at universities and a few leading defense contractors.

0 0

Post a comment