Simply, this experiment is meant to find the absolute upper limit for current Intel Pentium 4 Northwood processor speeds. This information is relevant in finding a balance between aluminum's conductivity, and silicon's semiconductor properties in heat efficiency. This top limit is acquired through a rather complicated process called overclocking. By manipulating certain values, one can overclock a computer. A good deal of care must be taken, however, because high processor load can cause visual corruption, booting failures, and more importantly, it can burn the processor out. To counterbalance this heat, overclockers will use heat sinks, fans, powerful water pumps, Thermo Electric Cooling plates (relying on the Peltier effect), ice, nonconductive liquids, and liquid gases. Such is the case with this experiment; it uses a large heat sink, fans, and above all, liquid nitrogen. Liquid nitrogen proved surprisingly effective for attaining extremely high clock speeds. It is an extremely effective, and equally unstable method of cooling. The highest clock speed reached was just under 3.9 Ghz. The highest 3Dmark bench was 16786 marks.
The objective of this experiment is, at its root, to build the fastest single-processor* computer possible out of available parts. The experiment is not entirely reliant on quality equipment, however. Such fantastic speeds can be achieved through a process called overclocking. The term "overclocking" refers to the crystal oscillator present in every computer. Its "ticks" are used to synchronize the execution of commands in a processor. Overclocking makes the clock's "ticks," the speed of which is measured in hertz, faster. As such, commands are executed more quickly, and the computer runs faster. Overclocking a computer is not as simple as that, however. The warranties of many products are nullified by overclocking, and with good reason: increasing the speed of a chip creates a large amount of heat, and with heat comes instability. Overheated computers will shut down spontaneously, fail to boot, display onscreen corruption of the video signal, and under very stressed conditions they may burn out completely.
To run such an expedient computer requires extremely low temperatures. There are a number of ways for a computer to be cooled: fans (in excess), heat sinks, Thermo Electric Cooling devices (relying on the Peltier effect), large water pumps in conjunction with heat sinks, Freon releasing systems (coupled with dehumidifiers), complete submersion in inert liquid, and contact with liquid gases. The last method, the use of liquid gas, is the one employed in this experiment. Liquid Nitrogen (LN2), at -196 degrees Celsius, is the coldest liquid gas that is feasible for use, both for expense and safety reasons. At such low temperatures, silicon is not as effective a semiconductor as it is under normal conditions. In order for the processor's silicon logic gates to function normally, it needs a stronger signal. As such, as the temperature decreases, it is necessary to increase the voltage supplied to the processor. While the silicon in the processor is adversely affected, the aluminum wiring's level of resistance is greatly reduced, making it a more efficient carrier for electrical signals. At very low temperatures, the balancing act that the silicon and aluminum are in becomes apparent. Past a point, the silicon is too stressed to function properly, and a large performance decrease can be noted. Ideally, in this experiment, the apex of performance can be achieved, at a temperature low enough to counterbalance the immense heat being emitted, but not so low as to make the processor function improperly.
Such experiments have been performed before. Holicho, a Japanese overclocker, has taken processors upwards of 4.5 Ghz with nitrogen cooling. He uses a large copper heatsink into which LN2 is poured. A team of Finns who run Muropaketti.com have used LN2 cooling on many occasions, and hold very high 3dmark scores. One of the more interesting nitrogen overclocking experiments was performed by a team representing OCTools.com. This was an experiment using an almost entirely inert liquid, Fluorinert. The motherboard was submerged in the liquid and then cooled with liquid nitrogen. The nitrogen was pumped through rubber tubing into an aluminum waterblock mounted on the processor, a modest Pentium 2 Celeron 556MHZ. It is important to note that the motherboard was first placed in a common freezer, to prevent the ceramic on it from cracking. Dry ice was used as well after it was submerged in order the cool the rest of the components more thoroughly. It should be noted that dry ice goes directly from a solid to a gas because of sublimation, so it would not short circuit any of the components as regular ice would. The liquid nitrogen soon caused the Fluorinert to freeze, and at that point the system became completely unstable. It is possible that the CMOS's battery was to blame, especially if the battery's acid froze.
*Dual processors are simulated with HyperThreadingFirst, the computer parts were cooled. The processor and motherboard were first chilled in a freezer to prevent cracks under extreme temperatures. A complete list of parts is given in the appendix, the most notable of which are the Pentium 4 3060 MHz Hyper Threading (Northwood) processor and the Abit IT7 Max2 Version2 (Intel 845PE chipset) motherboard. The waterblock heat sink, a Swiftech MCW5000-P, was mounted using Antec Reference Silver Thermal Compound thermal epoxy. The computer was booted and Windows XP Pro (for Hyper Threading support), was installed. The hard drive (or one small partition on it), an IBM 34.18GB SCSI Ultra160 10,000RPM drive, was defragmented with a program called O&O Defrag. There were a few steps to overclocking the computer.
1 Pour LN2 into the ½ inch tube in the waterblockIn addition to these steps, some other changes were made in the BIOS, namely RAM latency timings, for benchmarking purposes. The stipulations in step number 6 are made for practicality. After all, the computer is useless if it cannot sustain an operating system. A large mount was built out of spare parts to hold the pouring tube upright.
With enough effort, three fully successful overclocked boots were made. On every other attempt, the system crashed just upon loading Windows, or during a benchmark. The maximum attained speed was 3.89 GHz at roughly -70 degrees C. The estimate is based on the amount of time it took to get back up to -40 degrees C, which was the lowest the temperature sensor could accurately detect. The underclocked boots were made in order to test the processor out, and for heat reasons.
| 36 degrees C |
| 1894 MHz (Underclocked) |
| FSB: 100MHz |
| Score: 7786 |
| 30 degrees C |
| 1894 MHz (Underclocked) |
| FSB: 100 MHz |
| Score: 10462 |
| 25 degrees C |
| 3070 MHz |
| FSB: 133 MHz |
| Score: 13249 |
| -10 degrees C |
| 3466 MHz |
| FSB: 150 MHz |
| Score: 15737 |
| -70 degrees C (Estimated) |
| 3895 MHz |
| FSB: 169 MHz |
| Score: 16786 |
| -70 degrees C (Estimated) |
| 4140 MHz |
| FSB: 180MHz |
| Speed attained for insufficient time |
The results in the experiment were surprising in a number of ways. Though the experiment went well overall, there were many problems along the way.
The processor required a much higher voltage at very low temperatures than at first expected. At the lowest temperatures reached, no amount of coaxing would cause the computer to boot. The voltage that the motherboard supplied -- up to 1.8 volts -- was not sufficient. A solution should be sought for future experimentation.
Temperature monitoring was another difficulty encountered. In trying to forgo the (great) expense of an extremely low temperature thermometer, the processor's internal temperature sensor was relied on. After going beyond -40 degrees celsius, the sensor no longer gave readings. Instead, the temperature was very loosely judged by the amount of time it took for the processor temperature to get back up to -40 degrees.
Much more nitrogen was used than estimated beforehand. A gallon went by very quickly, because the waterblock provided a very gradual decline in temperature when nitrogen was poured in. As such, more had to be used to get the core processor temperature down to reasonable levels.
The highest. 3dmark score was fairly low, when compared to similar benchmarks submitted by others. While the processor clock speed attained was one of the highest ones submitted, (in 21st place), the 3Dmark score was the lowest of comparable benches. Much more weight (proportionally), it seems, is placed on the video card clock speed than on the processor speed.
The RAM often performed poorly in benchmarks and in general. It was frequently the cause of boot failure, and incomplete POST's. It was exceedingly troublesome when overclocked, especially when the RAM's latency settings were adjusted. A few page faults resulted as well. Because of this, in a future experiment ECC RAM may be used instead.
Windows was more susceptible to errors than anticipated. When the processor was slightly overheated, (even while underclocked), performance dropped very low, and many Stop Errors and Page Faults resulted. Just to boot Windows the waterblock had to be cool to the touch.
The video card became troublesome at times. When the PCI bus was accidentally overclocked, (because of the PCI bus speed's dependency on the FSB), the video card became unusable, as the AGP bus speed is simply the PCI bus speed multiplied by two. This also caused the SCSI card to cease functioning, not surprisingly. The parts were not damaged fortunately.
The means of cooling used was very different from those used by other notable overclockers. Both Holicho and the Finns (Muropaketti) used a very simple cooling process, where instead of a waterblock a square copper tube was used. The nitrogen in those experiments did not flow, as it did in this one. It may be the flowing nitrogen that resulted in inefficient cooling. As well, the vertical design used in Holicho's and Muropaketti's cooling setup enabled the use of insulation. OCTools had a similar setup for pouring in the nitrogen, but they used Flourinert in addition to the pipe and waterblock setup used in this experiment.
Thanks go to all of those who helped out in my project. Thanks to Stan G. for securing a gallon of LN2 from George Washington University. Thanks go to the Solid State Lab branch of GWU, who supplied the gallon of LN2 for free. Thanks to Swiftech for shipping the waterblock so quickly. Thanks to Newegg.com for shipping the processor so quickly, (and cheaply). Finally, thanks to McPhee.com for supplying the furnishings in the documentary, (giant hand chair).
"When the Intel Pentium 4 2,8GHz CPU arrived to our test lab we ordered 10 liters of Liquid Nitrogen (LN2 -196°C) and decided to run some tests in very low temperatures. After some adjusting and testing we were able to run SiSoft Sandra CPU and Memory benchmarks and Pifast benchmark smoothly when the CPU was running at 3917MHz. We raised the FSB one more step and managed to run successfully SuperPi benchmark while CPU was running at 3998MHz. The result was 39 seconds. Test setup: P4 2,8GHz, Modified Asus P4T533-C, Samsung PC800 RDRAM, PNY GeForce 4 MX440 and Windows XP OS. "
An experiment using an almost entirely inert liquid, Fluorinert. The motherboard was submerged in the liquid and then cooled with liquid nitrogen that was pumped through rubber tubing into an aluminum cooling block mounted on the processor, a modest Pentium 2 Celeron 556MHZ. It is important to note that the motherboard was first placed in a common freezer, to prevent the ceramic on it from cracking. Dry ice was used as well after it was submerged in order the cool the rest of the components more thoroughly. It should be noted that dry ice goes directly from a solid to a gas because of sublimation, so it would not short circuit any of the components as regular ice would. The liquid nitrogen soon caused the Fluorinert to freeze, and at that point the system became completely unstable. It is possible that the CMOS's battery was to blame, especially if the battery acid froze.
Not much to be said about this one. My lack of a working knowledge in Japanese is to blame for that.
More from OCTools. This is the precursor to "Mission Submersive 2", and goes into a bit more depth about the cooling process used.
"Liquid nitrogen can cool Athlon of a terrible calorific value to the limit." This site has a number of good pictures, and clearly, even better English.
More from those Finns. An earlier project with a 2.2 GHZ P4 Northwood, on an 845 DDR board from Asus. They used a Vcore hack for this, in order to get the processor to a sufficient voltage. As such, they ran the processor at an extremely high 2.12 volts. The highest stable clock they reached was 3.63 GHz. This page has a better English summary than the later one.
The aforementioned Vcore hack.
Pictures and statistics from a liquid nitrogen experiment with an Athlon T-Bird, a processor that runs at very hot temperatures even at regular settings. DDR speeds are very impressive, even when compared against current motherboards