Choosing the Best DDR3 SDRAM for Ivy Bridge

Processor microarchitectures continue developing, and the DDR3 SDRAM frequencies continue growing. However, does it really make sense to use high-speed memory with contemporary Ivy Bridge processors? To answer this question we analyzed the influence of memory frequency and latencies on the performance of a contemporary LGA 1155 platform.

by Ilya Gavrichenkov
07/06/2012 | 01:26 PM

It’s not easy to write something interesting about system memory because we have all got used to the fact that a computer's performance doesn’t depend much on the clock rate and timings of DDR3 SDRAM. Moreover, considering the low memory prices we have today, it is more reasonable for most users to invest into a larger amount of system memory rather than into improving its specs. The amount of memory is a much more tangible and comprehensible thing than its clock rate and timings whose effect is far from obvious.

 

That's why overclocker-friendly memory has become a product for perfectionists while average users have come to be satisfied with regular DDR3-1333 and DDR3-1600 SDRAM which they can occasionally overclock a little, just for fun. That’s the way things were and are, but how will they be? We raise this question because the new Ivy Bridge processor family has come about and, even though alike to their predecessors in many ways, the new CPUs have certain features that can change this situation.

For example, the Ivy Bridge series lacks the old limitations concerning the top clock rate of DDR3 SDRAM. In practical terms it means that such CPUs can theoretically be used together with high-speed memory up to DDR3-3200 SDRAM which doesn’t even exist as yet. The makers of overclocker-friendly memory modules have responded to this enthusiastically. Fast products, such as DDR3-2400, abound in shops now, so the gap between the clock rates of regular and premium memory products has got as wide as 100%! One can hardly believe that this difference can never show up in everyday tasks.

Here’s another argument in favor of this point: the Ivy Bridge series are generally faster than their predecessors and need to be fed data at a faster rate. In other words, it is quite possible that the new CPUs won’t be satisfied with the speed of DDR3-1333 and DDR-1600 in everyday applications anymore. The new integrated graphics core which shares the same memory subsystem with the execution cores must also be taken into account. The GPU has got faster in the Ivy Bridge design whereas memory bandwidth is a highly important parameter for today’s GPUs, directly affecting their texture-mapping speed. We guess the performance of the integrated graphics core is going to depend on the clock rate and timings of DDR3 SDRAM.

So, there are quite a lot of arguments in favor of checking out Ivy Bridge platforms with different memory settings. Our benchmarking results may turn out to be not as predictable as we think after all!

Testbed Configuration

To write this review we took an LGA1155 mainboard with Intel Z77 Express chipset and installed overclocker-friendly Core i5 processors, of both Ivy Bridge and Sandy Bridge generations. But since we wanted to see how memory subsystem settings affected our system performance, the key component was a high-speed DDR3-2600 memory kit from G.Skill.

Overall we ended up using the following hardware and software components:

Take note that we overclocked our CPUs to 4.5 GHz during this test session because this helped us get a better idea of the correlation between memory settings and performance.

Ivy Bridge Memory Controller

Memory controllers integrated into Intel CPUs have gone a long and winding way but the engineers seem to have finally found the optimal solution, so the Ivy Bridge memory controller has no significant differences from its predecessor. It is still based on an internal ring bus introduced back in the Sandy Bridge microarchitecture. The bus provides each execution and graphics core a quick and equal access both to the L3 cache and system memory. As a result, the peak data-transfer rate increases considerably and the Core CPUs for LGA1155 platforms are faster than their opponents in memory subsystem benchmarks.

So, it is no wonder that the same principle of interaction between the CPU and memory controller is implemented in the Ivy Bridge microarchitecture. Moreover, the engineers didn’t even try to revise anything in the controller’s internals. There are but minor improvements, especially as the Ivy Bridge series are perfectly compatible with the existing LGA1155 platform. It is still a dual-channel controller for DDR3 SDRAM but it supports higher clock rates than its Sandy Bridge predecessor: DDR3-1333 and DDR3-1600. The XMP technology has been upgraded to version 1.3. Both improvements can hardly be taken seriously because they do not mean much in practical terms.

Like its predecessor, the Ivy Bridge memory controller can work in symmetry mode (when the amount, clock rate and timings of memory modules in both channels coincide) as well as in compatibility mode which is referred to as Intel Flex Memory Technology. The latter’s point is in dividing the whole memory array into two parts, basing on the modules' specs: one with symmetric access mode and another with asymmetric single-channel mode. As a result, LGA1155 systems can be equipped with different memory modules without a catastrophic performance hit.

Each controller channel can work with one or two DDR3 SDRAM modules, either single- or dual-sided, so the maximum amount of system memory supported by today's LGA1155 systems is 32 gigabytes. For users who need more memory, Intel offers the higher-class LGA2011 platform.

Everything we’ve said so far in this section of our review is but a description of standard properties of Intel’s memory controller which applies to both Ivy Bridge and Sandy Bridge series. The new CPU design does have something new in terms of the memory controller, though.

There are two such things we want to note here. First, the frequency multiplier has become more flexible. With Sandy Bridge, DDR3-2400 was the highest memory mode possible whereas Ivy Bridge CPUs can clock system memory at frequencies up to 3200 MHz. And second, an additional variable multiplier has been introduced into the memory frequency formula that allows changing the clock rate with a step of 200 MHz besides the conventional step of 266 MHz.

All of this makes memory clocking much more flexible than before. Considering that LGA1155 systems do not allow changing the base clock rate, there are quite a lot of memory frequency options available now. Here is the full list for DDR3 SDRAM on an LGA1155 platform with an Ivy Bridge CPU:

It must be noted that this screenshot was captured on a Z77-based mainboard with a Core i5-3570K. Mainboards with H series chipsets are not that flexible when it comes to clocking DDR3 SDRAM. They are limited to the standard values of DDR3-1333 and DDR3-1600. Another limitation concerns Ivy Bridge CPUs other than the overclocker-friendly K series. They can only increase the memory clock rate to 2400 MHz.

So, it is with Core i5-3570K and Core i7-3770K processors that you can get as much flexibility in memory configuring as possible. Our experiments suggest that high-speed memory modes are perfectly functional and do not even require any tricks like fine-tuning secondary voltages. For example, it only took a small (by a mere 50 millivolts) increase in memory controller voltage for our Ivy Bridge CPU to work faultlessly with DDR3-2667 SDRAM.

As a matter of fact, Intel emphasizes the fact that it's extremely easy to reach high memory clock rates now. And you can keep your system stable by changing two voltages only. These are VDDQ, which is applied directly to the modules, and VCCSA, which powers the system agent and memory controller. It is not recommended to increase the former above 1.65 volts to safeguard the CPU against damage or degradation. The latter voltage is 0.925 volts by default and you can increase it a little to make your system more stable at high DDR3 clock rates.

Thus, the innovations in the Ivy Bridge memory controller seem to be targeted at overclockers who prefer running their DDR3 SDRAM in nonstandard high-frequency mode. As for the controller’s operation in its default mode, Intel doesn’t promise any changes compared to Sandy Bridge.

Anyway, we carried out a practical test to compare the Sandy Bridge and Ivy Bridge memory controllers as they worked with regular dual-channel DDR3-1600 at standard timings of 9-9-9-27-1N. We used the same LGA1155 platform, changing the CPUs only. For the clock rate not to affect the performance of the integrated memory controllers, the 22nm and 32nm CPUs were both overclocked to the same frequency of 4.5 GHz. Every power-saving technology, and Turbo Boost too, was turned off. Considering that the frequency of 1600 MHz can be reached in two ways with Ivy Bridge CPUs (1600 MHz DDR = 100 MHz x 1.33 x 6 or 100 MHz x 1.00 x 8), we checked out both of them.

The table below shows the results we obtained:

First off, we can note that the Ivy Bridge memory controller’s performance does not depend on the additional multiplier. We get the same results irrespective of whether we use a step of 200 or 266 MHz to build the memory clock rate. The difference amounts to a tenth of percent, which may just as well be due to some measurement inaccuracies.

Comparing the memory controllers of the Sandy Bridge and Ivy Bridge CPUs, we can see that they don’t seem to be identical. While the memory bandwidth is almost the same, the practical latencies can differ by a few percent. And it’s absolutely impossible to predict which controller is going to be faster in a particular test, so we can conclude that neither the Sandy Bridge nor the Ivy Bridge CPU offers specific benefits in terms of system memory performance.

Clock Rate vs. Timings

Every time we try to choose the best memory for our computers, we have to face the choice between higher clock rates and lower timings. However, this time around we are going to avoid detailed tests of DDR3 SDRAM modules with different timings because memory timings have come to affect the overall performance much less in the newer platforms than clock rate does.

There are two reasons for that. First of all, at higher clock rates the minimum timings are anyway quite high, so the value you can add is relatively small. When you add a couple of cycles to a timing that was originally set at 3 or 4 (like with DDR2 SDRAM), you are going to observe a bigger effect than if that timing was 7 or 8 cycles by default (as is the case with DDR3 SDRAM): the latency grows by 50-70% in the first case and by a mere 25-30% in the second case. Thus, different sets of memory timings are closer to each other with today’s memory.

The second reason is the overall improvement in how the CPU communicates with memory. There are more data caching levels and the amount of cache memory has steadily been increasing. This masks the actual latency of system memory, shifting the focus to its bandwidth instead.

As a matter of fact, the makers of overclocker-friendly memory kits have already realized that there's no need for low timings with high-frequency DDR3 SDRAM. Low-latency kits have disappeared and it’s rather hard to find DDR3 SDRAM modules with latencies below 9 cycles. The number of products with high latencies and high clock rates is, on the contrary, on the rise.

Well, we don’t want to be pure theorists, so we carried out a practical test to compare identical PC configurations with a Core i5-3570K and with DDR3-1600 and DDR3-1867 SDRAM that had different timings.

The diagrams illustrate what we’ve written above. Increasing the memory clock rate by 266 MHz turns out to be more effective than lowering each timing by 3 or 4 cycles. DDR3-1867 with 9-9-9-27 timings turns out to have a better effective latency than DDR3-1600 with aggressive timings of 7-7-7-21. As for effective bandwidth, DDR3-1600 can’t match the higher-clocked alternative under any circumstances.

Thus, memory timings have indeed become a rather insignificant parameter for today’s computers. When choosing DDR3 SDRAM for an Ivy Bridge processor, you should consider its clock rate in the first place whereas a low CAS Latency and other timings are in fact unimportant. The same goes for tweaking and overclocking: you should first focus on increasing the clock rate of your DDR3 SDRAM and only then minimize its latencies if you want to.

Clock Rate’s Effect on Performance

Now we’ve reached the main part of this review. We are going to find out how memory subsystem parameters can affect the overall performance of a computer in everyday applications. Considering what we've said in the previous section, we won't benchmark memory subsystems with different timings. Instead, we will focus on investigating the performance-frequency correlation. So we've taken popular memory configurations with clock rates from 1333 to 2667 MHz and selected typical timings for them. So, we ended up having the following DDr3 SDRAM types participate in our test session:

Otherwise, our testbed with a quad-core Core i5-3570K (Ivy Bridge) overclocked to 4.5 GHz was unchanged.

Synthetic benchmarks come first.

A higher clock rate expectedly helps to increase the effective bandwidth and reduce the effective latency of DDR3 SDRAM. Interestingly, we can see the largest increase in memory speed when the clock rate grows up to 2133 MHz. After that, the high clock rate doesn't have such a high effect. We can also note that switching to 1600 MHz has the biggest effect, indicating that DDR3-1333 looks like an outdated solution now. Overall, the 100% increase in frequency, from 1333 to 2666 MHz, leads to an up to 50% increase in effective bandwidth and latency.

The memory subsystem benchmark from Aida64 is single-threaded, so it doesn’t reveal the full potential of today’s memory controllers. Therefore we additionally ran the Stream benchmark in quad-threaded mode (to match the number of CPU cores in our testbed).

Indeed, the correlation between the memory subsystem’s bandwidth and clock rate is outlined sharper here than in Aida64. Switching from 2133 to 2400 MHz produces a noticeable effect, but the next 266MHz step doesn't seem so useful. Overall, when we progress from DDR3-1333 to DDR3-2400 or DDR3-2666, we can enjoy a 64% increase in data processing speed.

Well, these synthetic benchmarks can only draw some ideal picture but can hardly give us a notion of how the memory subsystem is going to behave in real-life applications. Let’s check them out, too.

Futuremark's benchmarks are not as enthusiastic as the synthetic ones: the clock rate does affect system performance but not that much. Increasing the clock rate by 266 MHz raises the overall score in PCMark 7 and 3DMark 11 by less than 1 percent, the physics test of the graphics benchmark being the only one to show some susceptibility to memory subsystem settings. That test runs up to 14% faster with overclocked DDR3 SDRAM.

It is in WinRAR that we can observe the biggest performance benefits from higher memory clock rates. In the rest of the applications fast DDR3 SDRAM can only speed the computer up by just a few percent.

We must admit that gaming applications differ somewhat in this respect. Memory subsystem performance has a larger effect on them. By preferring high-bandwidth DDR3 SDRAM for your Ivy Bridge platform, you can get an additional 5-10% in terms of frame rate. You don’t always achieve this even by installing a faster CPU!

Memory Clock Rate and Integrated Graphics Core

Graphics cores integrated into modern CPUs share system memory with the CPU proper. Therefore their performance is going to depend on the speed of DDR3 SDRAM installed. Moreover, considering that 3D rendering involves a lot of texture data transfers, this correlation may be even stronger than in conventional computing tasks. So, we want to see how the Intel HD Graphics 4000 core performs with system memory clocked at different frequencies.

Of course, we start out with the popular 3DMark 11.

Curiously, 3DMark 11 doesn’t think that the computer’s graphics performance depends on the memory subsystem bandwidth. The difference between the best (DDR3-2666 SDRAM) and worst (DDR3-1333) results is a mere 2.5%, which is even lower than in most non-graphics applications. The Ivy Bridge microarchitecture in which the graphics core has a dedicated cache and can also access the CPU's L3 cache seems to mask the effect of higher memory bandwidth on the performance of the HD Graphics 4000/5200 core.

The picture is somewhat different in actual games, though.

It looks like games do run faster on the integrated Ivy Bridge graphics when the memory subsystem has a high clock rate. Each extra 266 MHz of clock rate translates into a few percent addition to the frame rate, up to a total of 25%! It doesn’t mean the integrated graphics core itself gets faster because, as we’ve seen earlier, the frame rate goes up with a discrete graphics card as well. But anyway, if you plan to use the integrated graphics core of your Ivy Bridge processor, you may want to make it faster by installing high-frequency DDR3 SDRAM.

Closer Look at Our Kit of Choice: G.Skill [TridentX] F3-2600C10D-8GTXD

Our tests suggest that you should use high-frequency memory to achieve maximum performance on a platform with an overclocked Ivy Bridge processor. And when you go shopping for such memory, you will find that there are but few makers who offer products rated for a clock rate above 2400 MHz. Among them is G.Skill which was one of the first to provide dual-channel DDR3-2600 kits. That’s why we use their product for this test session. It’s the G.Skill [TridentX] F3-2600C10D-8GTXD kit consisting of two 4-gigabyte sticks. This memory is rated for 2600 MHz with default timings of 10-12-12-31-2N but, as we could make sure during our tests, it can do even better than specified.

Here is a summary of the product specs:

The memory modules are covered from both sides with red-and-black aluminum heatsinks called TridentX. Included into the kit is a frame with two 50mm fans you can install on the DIMM slots to cool this memory.

The active cooling is rather a marketing feature because, as our experience suggests, there is no real need for it. The modules do not get very hot at work.

What we want to make a special mention of is the shape of the heatsinks. As opposed to many other makers, G.Skill has addressed users’ complaints that tall heatsinks make memory modules incompatible with large CPU coolers. Therefore the new TridentX heatsinks are composite. The top (red) part can be easily detached after unfastening two screws to reduce the height of the module to a mere 39 millimeters.

The G.Skill [TridentX] F3-2600C10D-8GTXD kit supports XMP 1.3 for easy installation and configuring. The single predefined XMP profile, which is written into the modules in two copies, contains the specified clock rate and timings. Considering how easy it is to configure the Ivy Bridge memory controller, there's no difficulty in making this memory run at 2600 MHz. Just plug them in and get to work. There’s also a DDR3-1600 configuration with timings of 11-11-11-28 written into the modules’ SPD for the sake of compatibility.

The modules are based on Samsung’s popular K4B2G0846D-HCH9 chips manufactured on 30nm tech process.

These chips enjoy a good reputation thanks to their excellent overclocking potential and low heat dissipation. It is quite natural for them to be used in high-speed modules.


Makers of overclocker-friendly memory do not like to reveal the origin of their chips,
so the markings of off-the-shelf G.Skill TridenX series are going to look like this.

Our G.Skill [TridentX] F3-2600C10D-8GTXD kit could work not only at its default 2600 but also at 2666 MHz. We didn’t even have to change its timings for that.

Unfortunately, the next frequency of 2800 MHz could not be conquered at any settings.

The G.Skill TridentX series is designed for Ivy Bridge processors and Z77-based mainboards as such platforms can clock system memory at 2600 MHz and higher. There is a compatibility list for the G.Skill [TridentX] F3-2600C10D-8GTXD kit compiled by the manufacturer. So, this kit is guaranteed to work at 2600 MHz on the following mainboards: ASUS Maximus V GENE, ASUS P8Z77-V DELUXE, ASUS P8Z77-V PRO, Gigabyte GA-Z77A-UD5H, Gigabyte G1.Sniper 3, MSI Z77A-GD65 and MSI Z77A-G45. Of course, this doesn’t mean that you can’t try using this memory with other mainboards.

Conclusion

Our investigation of the correlation between the performance of Ivy Bridge platforms and their memory subsystem parameters suggests a lack of significant differences from Intel’s earlier platforms. The Ivy Bridge memory controller is largely the same as the Sandy Bridge one and delivers similar performance at the same settings. So, the influence of system memory settings on practical tasks is rather low. However, the new CPUs have brought about certain changes, the most important of which is the opportunity to choose a very high clock rate for DDR3 SDRAM. Such clock rates were not possible even with overclocker-targeted systems of the previous generation. As a result, the range of DDR3 SDRAM offered for LGA1155 systems has been extended, increasing the gap between configurations with slow and fast memory. By changing the memory clock rate alone, you can see a performance boost of 5-10% while applications that need large amounts of data (such as games) may get up to 20-30% faster! So, choosing the right kind of memory for you LGA1155 platform is important. We must note, however, that such benefits can only be achieved after a twofold increase in clock rate whereas a single 266MHz step up leads to a mere 2-3% increase in speed.

So, the rational approach to choosing system memory is in looking for the optimal frequency/price ratio. Modules up to DDR3-2133 SDRAM may be recommended for Ivy Bridge platforms, but not faster ones. Up to that frequency, the price remains reasonable while the performance in everyday applications grows up. Faster memory modules are considerably more expensive without providing tangible performance benefits, so they can only be recommended for enthusiasts who want maximum speed no matter what.

You shouldn’t bother about memory with low latencies, by the way. There are fewer such products on the market nowadays because low timings do not offer much in terms of performance on the modern LGA1155 platform.