by Ilya Gavrichenkov
08/09/2010 | 03:08 PM
After our investigation of the effect that memory frequency and timings have on performance of different LGA1156 processors, the next step is to see how things are with other platforms. So our next subject is the Socket AM3 platform which has become increasingly popular after AMD’s launch of good and inexpensive six-core Phenom II X6 processors.
As you probably know, the memory controller of AMD’s modern 45nm processors, including the Phenom II X6 series, supports both modern DDR3 and outdated DDR2 SDRAM. This ensures compatibility with older Socket AM2+ computers but it is obvious that such processors are only going to show their best on new platforms. When we got Phenom II X4 processors compatible with both memory types, we performed comparative tests and found out that new mainboards with DDR3 memory instead of Socket AM2+ mainboards with DDR2 provided a performance boost of 5 to 10%. However, DDR3 modules were considerably more expensive, and the choice was not so easy. By today, prices of DDR2 and DDR3 have become similar (DDR3 is even occasionally cheaper), so there is no reason to consider the slower DDR2 SDRAM now.
As for DDR3, the number of offers is simply mindboggling. Memory makers turn out a lot of products with different speeds, latencies and capacities, making it rather hard to find the most optimal one. So, we undertook a special test of different DDR3 SDRAM types to clear out the picture. Another reason for our test was the release of the latest generation of mainboards based on the AMD 890FX chipset. According to AMD and some mainboard makers, these products can finally support high-speed DDR3-2000 SDRAM (in overclocked mode, of course, but earlier products were explicitly incompatible with such memory). So, we wanted to check out such marketing claims and you’ll learn the results in this review.
We intend to study the dependence of the latest Socket AM3 platforms performance on the memory installed using the following testbed. It is built around a six-core Phenom II X6 processor and an AMD 890FX based mainboard from one of the leading makers. As a result, our test system was configured as follows:
Following our traditional test method, we first wanted to check out our system in its default operation mode, i.e. when no component was overclocked. We only changed the memory frequency multiplier and memory latencies. We tried to simulate normal working conditions during this test, so we did not turn off any of CPU technologies. AMD Turbo Core and Cool’n’Quiet worked as expected: our Phenom II X6 1090T processor sped up automatically to 3.6 GHz when only three or fewer of its cores had any work to do. When under higher load, the CPU worked at its default frequency of 3.2 GHz.
By the way, talking about memory tests, we want to remind you of what AMD has done to optimize software support for its platform. The AMD OverDrive utility makes it simpler to select and test optimal memory configurations on Socket AM3 platforms because it allows changing all memory controller parameters (and also some other settings) right from the OS, without requiring a reboot.
This utility also supports Black Edition Memory Profile which is similar to Intel’s XMP technology. AMD’s implementation does not provide for storing memory profiles in the module’s SPD. Instead, they are supplied with AMD OverDrive. The concept is interesting but hardly usable because BEMP supports but a few and rather rare memory modules as yet.
AMD made its dual-, triple-, quad-, and six-core Phenom II series processors compatible with DDR3 SDRAM clocked at 1067, 1333 and 1600 MHz. The Phenom II series is similar to Core i7 in this respect whereas the cheaper Athlon II, like the Core i5 series, doesn’t support DDR3-1600. We used a newest six-core Phenom II in our tests, so we could benchmark our test system with DDR3-1600 without any overclocking.
We didn’t have any problems with the Phenom II X6’s memory controller in default mode. The system could work at any memory speeds and timings and we did not observe any inexplicable performance slumps. By the way, unlike Intel processors with integrated memory controller, AMD processors do not put forth some special requirements about memory voltage. Increasing that voltage is not dangerous for the CPU. This might be expected since the integrated controller is compatible with both DDR3 and DDR2 although the latter type works at a higher voltage.
Now let’s see what results we’ve got. First of all, we ran synthetic benchmarks from Lavalys Everest.
Quite expectedly, the system memory delivers higher and higher performance as its clock rate grows up and its timings lower. A 266MHz increase in the frequency of dual-channel DDR3 leads to a 6% boost in terms of practical bandwidth at reading. Reducing the timings by one cycle lowers the latency by 4%. However, we shouldn’t forget that today’s CPUs feature a rather large amount of L3 cache which can smooth out the difference in memory parameters in most applications. Therefore, our final opinion should be based on numbers obtained in real-life benchmarks and algorithms.
We guess the diagrams don’t need much commenting upon. It is clear that memory settings have but a small effect on system performance. In fact, you can only feel the benefits of faster memory with reduced latencies in certain applications that deal with huge amounts of data, e.g. data compression or modern 3D games. So, it is only serious gamers that should indeed care about choosing the fastest memory modules available. Contrary to the widespread opinion, professional applications for digital content processing and creation are generally indifferent to memory subsystem parameters.
In the default mode the speed and timings of system memory have the same influence on the Socket AM3 platform as on the speed of Intel platforms we had tested earlier. Perhaps we’ll see some difference at overclocking?
Our Phenom II X6 1090T comes from the Black Edition series and thus can be overclocked via both the clock generator and the frequency multiplier. The second method doesn’t affect the memory controller, so we are not interested in it for now. Instead, we sped up our CPU for the second part of our tests by raising the base clock rate, leaving the frequency multiplier intact at 16x. We reached a frequency of 4.0 GHz, which is the clock rate you are likely to get with most Phenom II processors using air cooling.
Increasing the base clock rate raises the rest of system frequencies set up by means of multipliers. So, when the base clock rate is increased from 200 to 250 MHz, the processor begins to support another set of memory frequencies namely DDR3-1000, DDR3-1333, DDR3-1667 and DDR3-2000. Therefore, mainboard makers can claim that their products are compatible with memory faster than DDR3-1600.
Well, such marketing claims are sometimes too far-fetched. It is next to impossible to make system memory work in DDR3-2000 mode on a Socket AM3 platform. In fact, no one had ever mentioned such an opportunity before the release of the Phenom II X6 processor with the E0 stepping (this stepping features an optimized memory controller) and the AMD 890FX chipset. Then AMD said that this new CPU plus chipset combination could make DDR3-2000 possible if you’ve got a good mainboard, a lucky sample of the Phenom II X6 with an overclockable memory controller, and memory modules capable of working at such a high clock rate.
Mainboard makers were quick to spread out this news, omitting to note all the prerequisites, so some users may think that DDR3-2000 SDRAM is always an option with new mainboards if you’ve got a six-core CPU.
It’s not so in practice. Although there are accounts of successful experiments of this kind on the Web, they should be viewed as exceptions from the general rule. It is indeed a problem to make memory work at 2000 MHz on a Socket AM3 platform.
For example, we used an ASUS Crosshair IV Formula mainboard, presumably compatible with DDR3-2000, and a suitable Phenom II X6 1090T processor but could not make DDR3-2000 stable in our testbed. Therefore, we had to limit ourselves to DDR3-1000, DDR3-1333 and DDR3-1667 modes at a base clock rate of 250 MHz.
When the CPU is overclocked by increasing the base clock rate, the frequency of the CPU-integrated North Bridge (also referred to as Uncore; it includes L3 cache and a memory controller) may grow up along with the available memory frequencies. In our earlier overclocking guide we recommended to reduce the appropriate multiplier to keep the Uncore frequency as close as possible to the default 2.0 GHz. However, if you want to get the maximum from your configuration, you can try overclocking it as well. AMD hints at that, too. For example, the North Bridge of the server-oriented Istanbul processor, which is analogous to the desktop Thuban, is clocked at 2.2 GHz by default.
As we learned in our practical experiments, the North Bridge of our Phenom II X6 1090T processor could be easily overclocked by 50%, i.e. to 3.0 GHz, by changing the appropriate multiplier. Unfortunately, you can only increase this multiplier above the default 10x with Black Edition series processors. If you’ve got an ordinary CPU, you will have to put up with what overclocking you can get by increasing the base clock rate.
Does it make much difference, anyway? To check out the influence of the CPU-integrated North Bridge on system performance, we carried out a separate test of our Phenom II X6 1090T processor overclocked to 4.0 GHz with its North Bridge clocked at its default 2.0 GHz as well as at an overclocked 3.0 GHz. To keep the rest of the test conditions identical, the system memory worked in the same mode in both cases: DDR3-1333 with 9-9-9-27-1T timings. The next table shows the difference in performance we could observe in benchmarks and real-life applications:
The CPU-integrated North Bridge might be expected to affect system performance because it incorporates not only a memory controller but also L3 cache. Thus, increasing its operating frequency cannot but speed up the whole memory subsystem. On the other hand, the performance growth we can see in our tests is far smaller than the frequency growth. The 50% increase in frequency leads to but a 5% increase in performance, and not even in all applications. And if our CPU didn’t have unlocked multipliers, we’d see an even smaller performance growth while overclocking through the base clock rate.
In other words, there is an effect from changing the frequency of the Uncore part of a CPU but it is too small to affect the overall system performance much. You are likely to get much better results by simply overclocking your system memory. That’s what the next diagrams show. The results were obtained at an Uncore frequency of 3.0 GHz.
First, let’s check out the practical bandwidth and latency.
The results are roughly the same as in the previous section. Everest CacheMem seems to be indifferent to our overclocking of the CPU and memory controller, its results being almost the same as in the previous section of this review. However, there are notable changes which you can see in the test of copying data in memory. Overclocking makes the dependence of bandwidth at copying on the memory frequency more pronounced, so we can expect some changes in tests that are based on real-life applications and algorithms.
Well, these diagrams look much better than those that we had with the non-overclocked system. The overclocked processor calls for more data while the higher frequency of the CPU-integrated North Bridge lowers the latency of the CPU – memory thoroughfare. As a result, the frequency of DDR3 SDRAM begins to affect system performance even in such memory-indifferent tasks as final rendering and video encoding. You can get an average 5% performance boost by adjusting the memory parameters then. As for memory-sensitive applications, they can even speed up by 10-15%. This means that overclockers shouldn’t be as careless about choosing system memory as ordinary users.
As you may have noticed, we took high-quality overclocker-friendly DDR3 SDRAM memory kit of the Viper II Sector 5 series from Patriot for our tests, hoping that we might run our Socket AM3 platform with DDR3-2000 memory. But although these modules is certainly capable of working at 2 GHz (which is its rated frequency, actually), we could not use its capabilities fully because of unsolved problems with the AMD platform which, even in its newest implementation, proved to be not ready for high-speed memory.
In this section we will make our amends to the Patriot PVV34G2000LLKB modules and show that that was not their fault. You will see what this memory is really capable of and how the memory controller of the Phenom II X6 limits the frequency potential of DDR3 SDRAM.
So, the Patriot PVV34G2000LLKB kit consists of two 2GB modules of DDR3 SDRAM rated for DDR3-2000 mode with timings of 8-8-8-26. These modules are designed for LGA1156 systems, so their voltage is set at Intel’s recommended 1.65 volts.
The Patriot PVV34G2000LLKB kit comes from the Viper II Sector 5 series which means that these modules have large aluminum heatsinks. Well, there is nothing really extraordinary about the heatsinks which are black-anodized bars with comb-shaped top surface. They should be quite enough to ensure no overheat-related problems, though. You must be aware that the Patriot PVV34G2000LLKB, as any other tall memory modules, may have some installation problems if you’ve got a large CPU cooler.
The heatsinks are secured on the chips by means of gluey tape-like thermal interface. Below we found Elpida Hyper chips we had praised a number of times in our earlier DDR3 reviews.
After all that, it is no wonder that the Patriot PVV34G2000LLKB has the following official specs:
The kit also includes a registration key for the popular benchmarking suite Futuremark 3DMark Vantage.
Being targeted at LGA1156 platforms, these modules support XMP technology.
There are as many as three modes in the profiles: DDR3-2000 with 8-8-8-26 timings, DDR3-1750 with 7-7-7-23 timings, and DDR3-1500 with very aggressive timings of 6-6-6-20.
Thus, the Patriot PVV34G2000LLKB features very attractive specs but it can only show its best on LGA1156 platforms together with Core i7 processors. Using an ASUS P7P55D Premium mainboard (Intel P55 Chipset) and a Core i7-860, we could make this memory stable as DDR3-2119 with 8-8-8-24-2T timings. The memory voltage was 1.65 volts, in full compliance with the specs and Intel’s recommendations.
Thus, the Patriot PVV34G2000LLKB is excellent system memory for enthusiasts. It works not only at 2000 MHz with rather aggressive timings but can also be overclocked to much higher frequencies.
But as we wrote above, the Socket AM3 platform proved to be unable to clock system memory at such a high frequency, so we could only reach DDR3-1833 mode with 8-8-8-24-1T timings and 1.65V voltage using an ASUS Crosshair IV Formula mainboard with a Phenom II X6 processor.
Comparing this result to what we achieved with the Intel platform, we can be quite sure that the problem is not about the memory modules. It is the CPU’s memory controller and mainboard that are to blame as they cannot clock system memory at high frequencies despite our complying with all formal requirements to using DDR3-2000.
We did not succeed even when we selected relaxed memory timings or increased the DIMM voltage.
Even at 10-10-10-30-2T timings the highest memory frequency we could achieve was 1873 MHz. So, we cannot say as yet that the Socket AM3 platform supports DDR3-2000. You can only achieve that frequency with a Phenom II X6 processor if you carefully select appropriate components. In most cases, purchasing high-speed memory for the latest AMD platform brings no practical benefits.
We have not seen anything unexpected in our today’s tests. Like with Intel platforms, the frequency and timings of DDR3 SDRAM may have but a small effect on performance of the Socket AM3 platform. A modern processor has enough cache memory to smooth out any latency that occurs on the path that data takes towards the execution cores. Working in its nominal mode, our Socket AM3 system with a newest Phenom II X6 processor (which has 6 MB of L3 cache) would only get an average 2-3% faster at higher memory settings. The performance may increase up to 5% in some applications that operate with huge amounts of data but this doesn’t change the overall picture. When the system components were not overclocked, there was almost no difference as to what memory was installed: DDR3-1067 or DDR3-1333 or DDR3-1600. You only have to make sure that it works in dual-channel mode.
This rather boring situation can be enlivened somewhat by overclocking. AMD processors allow to overclock the memory controller and L3 cache separately from the execution cores. This overclocking is quite rewarding. Rated for 2.0 GHz, the Uncore part of the CPU can be clocked at much higher frequencies, which leads to a performance growth up to 5%. You shouldn’t neglect this if you really care about your computer’s memory performance.
When you increase the base clock rate, the frequencies of the execution cores and the Uncore part of the CPU grow up and higher DDR3 SDRAM frequencies become possible. Coupled with the increased gaps between the available memory frequencies, this makes the effect of different memory frequencies and timings easier to see. For example, the difference between the slowest and fastest memory mode at a base clock rate of 250 MHz amounts to an average 5%, reaching 8-9% in some applications. And we even could not start our system up using the fastest DDR3-2000 mode, which requires additional comments.
AMD and mainboard makers have assured users that new mainboards based on the AMD 890FX chipset support DDR3-2000 mode if used together with a Phenom II X6 series processor. However, this is just a marketing claim that is not well grounded. The sentence lacks an important reservation like “if the planets of the Solar system line up in a particular order” or something. During our tests we could only reach memory frequencies of 1800-1900 MHz which is the top limit you can expect from good hardware. Higher results are only possible if you carefully pick up the most overclocker-friendly samples of mainboards and, most importantly, processors.
Thus, most AMD users cannot benefit from high-speed memory available in shops today. They should not care to purchase overclocker-friendly modules with a rated frequency of higher than DDR3-1866 as they won’t be able to utilize the speed potential of such memory fully.