02/24/2009 | 07:20 PM
The variety of components for the new LGA1366 platform continues to shape up. Right after the official launch any mainstream user could purchase a new Intel Core i7 processor and a corresponding mainboard in order to put together a new platform just like that. However, overclocking fans and computer enthusiasts faced certain complications. First of all, there were very few mainboards to choose from, normally just one or two models – now the choice is becoming richer day by day. At first, it was very difficult to find an efficient processor cooling solution – now all the leading cooler makers are offering special LGA1366 retention kits for their existing high-end products and make all their new solutions compatible with the new platform right from the start.
Choosing the right memory for a platform like that also turned out a little challenging. The first problem is evident: one or two memory modules are no longer enough, because triple-channel memory access can only be implemented with three DIMMs in the system. And again, it was no big deal for a mainstream user. He or she could go and buy any three regular DDR3-1066 memory modules, because in LGA1366 platform they would work at 533MHz frequency in the nominal mode at 133MHz base frequency setting. Overclockers and enthusiasts prefer to work at higher memory frequencies, that is why there have long been special high-frequency DDR3 memory kits available to them. But there is one problem: to get these memory DIMMs overclocked to higher frequencies they needed to increase their voltage to 1.9-2.0V. However, it is not recommended to push the memory voltage beyond 1.65V on LGA1366 platforms, because of the memory controller integrated into Intel Core i7 processors. As a result, it turned out impossible to take full advantage of the “old” DDR3 overclocker kits in the new systems. Insufficient voltage wouldn’t let them reveal their full potential.
However, this problem has already been solved. All leading memory makers offer special memory kits for the new LGA1366 systems. They include three DIMMs that can work at high frequencies and only 1.65V voltage. Today we are going to discuss one of these kits. We will study its features and functionality and check how big of a performance boost it will grant us in the nominal as well as overclocked mode.
Of course, the kit looks very impressive due to tall heat-spreaders:
I have always believed that the tallest heat-spreaders belonged to Corsair Dominator memory modules. Now I realize that I was wrong: Kingston heat-spreaders are 1.5 times taller:
In fact, the memory module equipped with this heat-spreader is twice as tall as a standard memory module:
These heat-spreaders look very impressive, however, they may cause some serious difficulties during system assembly. They may prevent you from successfully installing some large CPU coolers. But you put up with whatever necessary to achieve the desired goal, which in this case is to get your memory working at high frequencies. However, the most frustrating thing about it is that all these sacrifices are absolutely unnecessary: you do not need these gigantic heat-spreaders at all. You can easily see that DDR2 Corsair Dominator memory does need tall heat-spreaders, because these modules work at higher 2.1V voltage and heat up a lot under heavy load. As for DDR3 memory from Kingston working at 1.65V, it remains cool in idle as well as burn modes. These tall heat-spreaders causing quite a bit of inconvenience are mostly created for aesthetic purposes, to make the users feel how reliable and powerful this memory solution is. There is very good proof of this assumption: Kingston also offers the same exact memory modules equipped with the traditional heat-spreaders but featuring identical specifications: they didn’t have to lower the voltage or increase the memory timings because they no longer had gigantic heat-spreaders.
Kingston Company uses very clear and easy to understand product indexing system for their memory modules. The product model name tells you all you need to know about the parameters of the solution:
Let’s decode the product model name on our memory modules - KHX14900D3T1K3/3GX:
Let’s use Everest to check out the memory modules SPD and the information about the timings and frequencies combinations stored in it.
To ensure that this DDR3 SDRAM will be compatible and will start without any problems at the nominal voltage settings of 1.5V, the modules are introduced to the mainboards as standard DDR3-1333 working at 667MHz with 9-9-9-24 timings. SPD also contains information about other possible frequency and timings combinations. You will have to set the default voltage of 1.65V, frequency of 1866MHz and timings of 9-9-9-27 on your own in the mainboard BIOS. However, if the board supports Intel XMP technology, all these parameters will be adjusted automatically once you select the corresponding profile.
We are going to test the performance of Kingston HyperX DDR3-1866 memory kit on the following testbed:
In the default mode Gigabyte GA-EX58-Extreme mainboard, like other Intel X58 Express based boards, sets the memory frequency at 533MHz, i.e. works with the memory as DDR3-1067. The timings in this case are set at 9-9-9-24-1T. The board supports Intel XMP technology that is why it adjusts the DDR3 settings above the nominal for the regular DDR3 SDRAM but to the standard values for Kingston HyperX DDR3-1866 once you select the first profile in the mainboard BIOS.
If you choose the second profile, the board starts acting in a very interesting way. This profile has the memory working at 900MHz, i.e. as DDR3-1800. In this case the board lowers the processor clock frequency multiplier from x20 to x17, but raises the base frequency from 133MHz to 150MHz. As a result, the CPU frequency is 2.55GHz, which is as close at it can get to the nominal 2.66GHz. The memory, however, works exactly at 900MHz, i.e. as DDR3-1800.
In both cases the board increases the memory voltage to 1.65V. However, I was a little confused to see that the QPI bus voltage rose to 1.5V in the first case and to 1.45V in the second, which is still way too much. I have to remind you at this point that the default QPI/VTT Voltage is only 1.175V.
I failed to get the system to start at 933MHz memory frequency after only raising the voltage to 1.65V. The board would boot but hang right during POST when it got to the memory part. Then I increased the QPI/VTT Voltage to 1.335V. I chose this number almost randomly: the next available setting of 1.355V was already considered dangerously high and was highlighted purple in the mainboard BIOS. It was a perfect guess: the system remained stable at this QPI/VTT Voltage and passed all tests, but we couldn’t set this parameter any lower than that. We needed to get to the lowest QPI/VTT Voltage setting possible, because the higher you set it, the more it increases the processor temperature. You can barely notice it in the nominal mode, but during overclocking the temperature may rise dangerously high.
As a result, we tested our Kingston HyperX DDR3-1866 memory modules at the voltage increased to 1.65V and QPI/VTT Voltage set at 1.335V. We used System Stability Tester utility calculating 2-8 million digits of Pi in multi-threaded mode. The system was considered stable if it could also pass a 1-hour Prime95 test in Blend mode. Unfortunately, we couldn’t increase the frequency beyond the nominal 933MHz at 9-9-9-24 timings. This is pretty strange because memory doesn’t usually work at the maximum of its capacity, just like other system components. However, it could be the Gigabyte GA-EX58-Extreme mainboard that we used. We are going to repeat our tests on a different mainboard shortly and will let you know what we find.
Memory modules SPD states that at CAS Latency of 8 the memory can work at 829MHz frequency as DDR3-1658. In reality our memory kit did much better and passed the tests at 906MHz, i.e. in DDR3-1812 mode. It should work at 725MHz (DDR3-1450) at CAS Latency 7, but in fact it worked as DDR3-1530. As for the CAS Latency 6 at 1.65V voltage there is no info about it in the memory modules SPD. All we know is that the frequency will be at least 444MHz at 1.5V voltage setting. The kit passed all tests at 695MHz, i.e. in DDR3-1390 mode.
The results are overall very good, especially keeping in mind that they were obtained I triple-channel mode and at a relatively low voltage setting of 1.65V. However, these are only preliminary data that were obtained in easy conditions with little or no CPU overclocking, i.e. in nominal processor mode. We used the following combinations of base and memory frequencies: 133/1867 MHz with 9-9-9-24 timings, 151/1812 MHz (8-8-8-22), 153/1530 MHz (7-7-7-20), and 139/1390 MHz (6-6-6-18). I wonder if we can get the same results during actual system overclocking.
I have first used Intel Core i7-920 processor during the tests of Gigabyte GA-EX58-UD5 and GA-EX58-Extreme mainboards. It could overclock up to 181MHz base frequency at its nominal .225V Vcore, i.e. with all power-saving technologies up and running. With the Vcore increased to 1.3V, it remained stable at 188MHz. In both cases the memory worked at its maximum possible frequency of 1067MHz and 1128MHz respectively. We also managed to lower the timings to 6-6-6-18, which is pretty aggressive for DDR3.
When we overclock our CPU to 181MHz base frequency, we can set the memory frequency at 1810MHz – the closest value to the nominal 1867MHz. The obtained results suggest that we can even lower the timings to 8-8-8-22 in this case. Although this frequency is dangerously close to 1812MHz, which is the maximum frequency for the memory to remain stable with these timings, the system passed the stability tests successfully. We just had to change the Performance Enhance parameter in the mainboard BIOS from Turbo to Standard.
When the CPU is overclocked to the maximum of 188MHz base frequency, you can set the memory speed at 1880MHz, however, this value is outside the interval supported by Kingston HyperX DDR3-1866. We did try to conquer this frequency by raising the voltage, but it led to CPU temperature increase beyond 90°C, instead of system stability. To ensure that the memory would work at its absolute maximum frequency, maybe even at the expense of the resulting CPU frequency, we lowered the base frequency to 186MHz. The memory frequency in this case equaled 1860MHz, but even in this case the system remained unstable.
As a result, we had to set the memory frequency one step lower to 1504MHz. Theoretically, you could lower the memory timings to 7-7-7-20 in this case, which we did. By the way, even in this case when we set the processor Vcore to 1.3V, enabled Load-Line Calibration technology and set QPI/VTT Voltage to 1.335V the maximum CPU temperature was 88°C after one-hour run of Prime95 test. It is too high for a high-end Cooler Master GeminII cooler with a 120-mm fan working at 2500RPM and a system assembled in and open Antec Skeleton system case.
Tests have been completed and it is very interesting to see how much of a gain we get when we increase the memory frequency that dramatically? At first let’s compare the performance of a system working in nominal mode with the system where memory frequency is increased to 1867MHz. We’ll start with special memory benchmarks, such as Everest 5.0, BenchDLL version 2.4.258.0.
The performance boost is gigantic, between 30 and 52%! And what will we see if the performance depends not only on the memory subsystem, but also on the CPU or graphics card?
Here things a little more modest, I should say, and the minimal frames per second in Crysis Warhead are even considerably lower. However, these results vary too much during different test cycles. For example, during the first test cycle the minimal fps rate was only a little over 17fps in all ten runs. The table below shows the average of the second ten runs, and third time we would have most likely got a third number, so I wouldn’t take this parameter too seriously. However, the overall performance boost resulting from the increase in the memory frequency from 1067 to 1867MHz is not very breath-taking. Let’s check out the results during overclocking, to see if the performance gain is going to be more noticeable. We will start with Everest benchmarks.
The improvements are quite noticeable for the latency and read speed, while for the write speed and copy speed – not really. What will the tests in other applications show?
Unfortunately, the performance gain is mostly minimal, less than 1%, i.e. could be considered being within the measuring error. I have to admit that it is very disappointing. However, we did have to lower the memory frequency quite noticeably, even though the timings have been lowered accordingly. Maybe this is why we do not see a significant performance boost? Let’s repeat the tests with the CPU overclocked to 181MHz base frequency. In this case not only the memory frequency is close to the maximum, but the timings can be lowered from CL9 to CL8. It is a simply ideal situation to reveal the advantages of high-frequency memory:
The situation in Everest test reminds us of what we have just seen I the very first table: the system with high-frequency memory is way ahead. Let’s check out other applications now:
Unfortunately, we didn’t see any significant advantages. Like in the previous cases, the performance improvement is most evident in a multi-threaded Custom PC Bench test. During this test there is a video being played back while 7-zip archives some files in the background. It is a very real situation: say, you are watching a movie and the operating system starts backing up and archiving the files. It’s a pity that this is the only other type of scenario besides Everest benchmarks when we see an indisputable advantage of high-frequency memory.
Kingston HyperX DDR3-1866 memory modules have only one disappointing drawback: extremely tall and unnecessary heat-spreaders that may cause some problems during system assembly. High price of this kit may also have a little discouraging effect on you, however, it is similar to what over overclocker memory kits run at these days. You can currently order a kit like that on Kingston’s web-site for $237. A memory kit with exact same specifications and features but with regular heat-spreaders - KHX14900D3K3/3GX – is selling for $219.
In my personal opinion, it would be a better buy, although it is hard to justify the almost 10 times higher price than that of regular DDR3-1067. Of course, overclocker memory that can work in a much broader frequency range, like Kingston HyperX DDR3-1866, offers overclockers much more flexibility. You will always be able to find the most optimal frequency and timings combination. But you have to decide if the result you will get in a few selected applications sensitive to memory frequency and timings is worth the investment.