We let Sweden’s best CPU overlocker take on a Pentium 4 660 processor with most cooling technologies imaginable. Everything from air to double compressor coolers, the results can be found in our latest article.
Intel has found themselves in the wake among enthusiasts after the launch of the Prescott core that shocked everybody with its enormous heat dissipation and lack of increased performance. Intel thought that they could increase the clock frequencies and therefore gain what was lost by adding another pipeline-step. Since then, Intel’s engineers have worked hard with refining the manufacturing technique and to reduce the current leakage in order to increase the frequencies. The stepping is now at its third edition and despite the fact that it still have problems with the heat dissipation; they have managed to get the CPU up to higher frequencies.
Earlier we have made a similar review of the AMD FX-57 CPU, and even if we didn’t get hold of Intel’s fastest CPU for this test, this CPU should give us an indication of how the general overclocking capabilities are. As said, Intel is on its third big revision of the Prescott-kernel that has the name E0/N0. E0 is for the 500-series and the Celeron D CPUs while N0 is for the 600-series. N0 have 2MB “level 2”-cache while E0 only have 1MB (of which only 256KB is activated on the Celeron D-version). In addition to this, the N0-stepping also has support for EIST, which enables it to underclock itself when the system is not under load.
Let’s have a look at the test system and comment on the structure of this article.
 | |
Test system | |
| Hardware | |
| Motherboard | Asus P5WD2 Premium, BIOS 0606 |
| Processor | Intel 660, 3.6GHz (Prescott, 0442) |
| Memory | Mushkin High Performance PC4200 DDR2 |
| Video card | nVidia GeForce 6600GT |
| Power supply | OCZ PowerStream 520W |
| Software | |
| Operating system | Windows XP (SP2) |
| Drivers | Intel Chipset Driver 7.2.1.100 |
| Monitoring program | Asus AI Booster SysTool |
| Test program | SuperPi 1.1e |
We have chosen to run these tests on a motherboard from Asus, the P5WD2 Premium. This board is known among Intel-overclockers as the best on the market when it comes to overclocking. The two most interesting reasons worth mentioning: The first is that the mainboard is based on the Intel 955X-chipset; this chipset can reach the high FSB-frequencies that we require. The second reason is that Asus finally have started using a 4-phase power regulator, which is a great advantage when using the voltage consuming Prescott CPUs.
We will, as in the previous article, mainly concentrate on how much the CPU can be overclocked using different types of cooling-devices. To test the stability of the CPU we will be using the software SuperPi and even if the CPU doesn’t have to be very stabile to run a 1M calculation, it takes a lot more to run a 32M calculation.
The test will begin by looking into how the most common heatsink performs, Intel’s own.
The stock cooler that comes with the P4 CPUs varies depending on what factory manufactured the CPU and what model it is. The higher frequency models have a copper-core while the models with lower frequencies sometimes only have an aluminum-core. The heatsink that we will be use today is the copper-core version, which according to our tests, is a lot better than the one with the aluminium-core.
 |  |  |
Knowing that the Prescott-CPUs emit lots of heat and that there often is a margin concerning the standard-voltage, we decided to start with a low voltage, below the standard-voltage. The standard voltage is 1.4000v but we choose to begin at 1.350v.
The stock cooler performs its task well when it comes to the standard frequency, which in this case is 3.6GHz. When we started to overclock the CPU, the temperature started to rise above 60°C and we choose to end its pain and stop at 1.375v. This doesn’t mean that the CPU will become unstable when we use 1.4v, but that the increase from 1.350v to 1.400v will generate more heat in relation to the high frequency, which leads to worse overclocking. This is worth considering before classing the P4 as a bad overclocker – try to lower the voltage.
Now lets try an old favourite from our previous overcklocking article; The Thermaltake Tower 112 with two Delta-fans. To make it a lot worse, ergonomically, we had the windows wide open. This resulted in the fact that the air temperature going through the heatsink was 10°C.
 |  |  |
Thanks to this increase in cooling we could start to test different voltages. We passed the 1GHz overcklocking limit at 1.475v. At 4.6GHZ, AIBooster reported a full load CPU-temperature of 37°C, compared to 60°C for the stock-cooler. In order to pass a 32M calculation, that takes about 20 minutes, we hade to lower it to 4418MHz at 1.45vcore.
The cold definitely gives positive results, so let’s rig the water cooling and se how that works out.
We have previously reviewed the water cooling (check out the review here), which basically consists of an Asetek Antarctica water block and a radiator which uses three 120mm fans. As with the air cooling, we opened the windows, to decrease the water-temperature considerably. After a couple of minutes, the temperature stabilised at 10°C and we started our tests.
 |  |  |
The highest CPU-temperature we got from AIBooster using this cooling-device was 24°C, a considerable change compared to the aircooling. But what about the result? Considering the low temperatures, we had hoped for better numbers and better tolerance when it comes to higher voltages, but that wasn’t the case. The strong side of watercooling is just that, water and its large heat capacity and we had hopes for a better result from a 32M calculation. With an optimal voltage of 1.50v, we managed to run such a calculation at 4612MHz which must be considered acceptable.
We’ve done all we can with conventional cooling-devices and its time to bring out some more powerful equipment.
To increase the performance of a water cooled system you can use different methods, like Peltier elements or you can use imaginative methods, e.g. tap-water or fridges/freezers. But that’s nothing we will be writing about in this article; instead we will go directly to compressor cooling. You can buy commercial units from some online stores, but we will be using a homemade system with better cooling capacity. The device works just like a freezer, except that the cold is concentrated on small area, the evaporator that is attached to the CPU. This unit consumes about 600W and manages to keep the temperature at -40°C at a load of 150W. As you can see in the following pictures, the CPU-socket must be isolated in order to avoid condensation.
 |  |  |
As we suspected, this hot CPU likes cold. Already at 1.350v we pass the air- and watercooled solutions and at 1.425v we pass another milestone: 5GHz. The CPU continues to scale nicely until we reach 5200MHz, where it starts to become a bit harsh and despite the fact that the temperature is under control, higher voltages won’t work.
When we reached just over 5200MHz we had, during full load, a temperature of -38°C on the evaporator and a CPU temperature of -19°C. After testing the voltages, it showed that the maximum frequency where a 32M calculation could be performed without errors was 5016MHz at 1.550v. When it comes to this CPU, we’re starting to see hints of a trend, where the most important part is not to increase the voltage but to decrease the temperatures. Let’s move on to the next compressor cooling-device.
The compressor cooling device used in the above test is not common, but you can buy them. The device we’re going to use now is way beyond what you normally can find here and there. A cascade compressor cooling device is based on the connection of two or more compressor coolers in a series, which cools each others gases in steps, in order to achieve extremely low temperatures. Our unit is a two-step device with R404A-gas in the high stage and R1150 gas in the low stage and can generate temperatures as low as -100°C.
 |  |  |
Cold is the ultimate cure and we notice the same phenomena that we noticed when we changed from water cooling to the one stage device, the same frequency as the previous device but at the lowest voltage. AIBooster didn’t manage to give us any reasonable temperatures, but our external temperature meter showed temperatures around -100°C during full load all the way up to 5.6GHz. The temperature rose to -98.7°C when we hit 5805MHz at 1.675v. Because we didn’t want the temperatures to go haywire during a 32M calculation, we had to go down to 5407MHz at 1.50v. Here are some performance numbers using these speeds.
 |  |
We move on and analyse the results on the next page.
Here are all of the results in the same chart. If we draw parallels between our experiences with the AMD FX-57, we see that the Intel 660 doesn’t scale as well when increasing the voltage, but a lot more when decreasing the temperature. When you use air cooling you have to be careful with the voltages, the heat emitted at the CPUs original speed is very high. We reached 60°C using the original heatsink despite the fact that the air temperature was about 22°C. Although the CPU didn’t start to throttle until about 70°C, the margin was too tight and an increase in the voltage didn’t result in any positive effects.
With some help from cool air, the bigger heatsink and the water cooling managed to tame the temperatures enough for us to be able to increase the voltage a few steps. Getting a 1GHz overclocking is nothing to snort about and neither is a 32M calculation at a stable frequency of 4.6GHz using watercooling. And when we lower the temperatures with about 50°C, we see a substantial rise in the frequencies and a better response when increasing the voltage. With another 50C temperature decrease, the scenario repeats itself. Unfortunately, our cascade device isn’t adjusted to the extreme effect the CPU emits at these speeds during longer periods. And therefore, we had to decrease it to 1.500v and 5400MHz for the 32M calculation. At 5600MHz and 1.575v, we could run half of the 32M-test before we had to abort in order to save the cooling device. The CPU uses roughly 200W at these speeds.
Let’s summarize our experiences of this CPU.
In our previous article about extreme overclocking, we established that more cold leads to better performance. This turns out to be even truer when it comes to the Intel P4 CPU, which scales a lot better using better cooling than when increasing the voltage compared to the AMD FX CPU. In the previous article, we also discussed the so-called cold bugs, which mean that the CPU won’t function properly at low temperatures. So far the Intel CPUs have been spared from these bugs and as the graph indicates, there is more to gain when using e.g. liquid nitrogen.
In the beginning of the article, more exactly the part about the stock cooler, we reflected over the fact that you shouldn’t increase the CPUs voltage too early. As we saw, the processor overclocked as high as with the original voltage as it did with 0.05v below it, with the benefit that the CPU didn’t get as hot. This could be something to think about for every overclocker using aircooling.
The degree of overclocking for a CPU depends on many different factors, where some of the most important are model and manufacturing week. The model is important considering the fact that the higher models are often the one that reached the furthest in the testing phase at the factory. This model is manufactured week 44 in 2004, thus relatively old, but it still overclock relatively good. We have previous experiences with a P4 630 (3GHz) with almost the same manufacturing date, that didn’t make it over 4GHz using watercooling. As time pass, the manufacturing process is adjusted and optimized and the lower models start to overclock better and better, which is a trend we’re starting to see in CPUs manufactured in and after the beginning of the year 2005.
We hope that we have given you an insight in how the other field of the CPU market looks like and what kind of overclocking you can expect from these frequency beasts. Finally, we would like to thank Intel whom has supplied us with this CPU.
