15 Jul 2012
Posted by synt4x 
Loading ...
Development of the processor microarchitecture does not stop, continue and progress in the frequencies of DDR3 SDRAM. But whether it makes sense to use high-speed memory with modern processors Ivy Bridge? To answer this question, we analyzed the effect of frequency and memory latency on the performance of a modern version of the LGA1155.
Investigation of the effect of high-speed memory settings on the overall performance of modern systems – employment, in general, the ungrateful. Such materials are often very popular among the readers, because everything is already used to the fact that the dependence of speed on the frequency and timing set out in the platform DDR3 SDRAM – very weak. Even under present conditions, when memory prices are through the floor, far more sensible strategy for most users seem more investment in increasing the total memory, but not in improving its performance. At the very least amount of memory – far more tangible and understandable value than its frequency and latency, whose connection with the responsiveness of the system is not so obvious.
All this has led to the fact that over time overclocker memory went to the category of goods for perfectionists, but the average users are satisfied with ordinary modules, DDR3-1333 or DDR3-1600 SDRAM, which at best will be a little broken up, they say, just for fun. So it was, so it is, but will continue, and so on? This question arises in connection with an appearance on the market a new family of processors Ivy Bridge, which let in many respects resemble their predecessors as two drops of water, but at the same time and have some features, making adjustments to the usual picture of the world.
For example, processors Ivy Bridge lifted old restrictions on the maximum frequency supported by the DDR3 SDRAM. In practice this means the theoretical possibility of operation of these processors with high-speed memory up to DDR3-3200 SDRAM, which, however, does not yet exist in nature. However, manufacturers of overclocker memory kits embraced this change the “rules of the game” with great enthusiasm: to find in stores, for example, DDR3-2400 modules are now not difficult. As a result, the gap between the frequencies of “running” and “elite” versions of the memory has reached up to twice the size. In fact, that it will not have a significant impact on the performance of systems commonly used tasks, at least not easy to believe.
There is one more argument: Ivy Bridge processors have become more productive than their predecessors, and consequently, their need for fast data processing has increased. In other words, it is possible that finally passed the critical threshold to which the rate is not a quick memory like DDR3-1333 or DDR3-1600 is quite enough for most needs of the processor. Especially as a significant contribution to overcoming this line could well make a new graphics engine, and using a common memory subsystem along with the cores. Performance graphics core in the Ivy Bridge jumped quite significantly , and memory bandwidth for today’s GPU – one of the fundamental characteristics, providing a direct influence on the rate of texturing. Therefore, it is reasonable to expect that, if not a traditional computer performance, it really speed integrated graphics will show pronounced dependence on the frequency and latency of installed modules, DDR3 SDRAM.
Either way, the argument for the study of performance-based systems Ivy Bridge, equipped with various DDR3-memory abound. Despite the prevailing skepticism about the predetermination of the results of this testing, we believe that the situation could change radically. And in confirmation of offer practical test results.
Description of test systems
In preparing this material was involved Platform LGA1155, built on a modern motherboard with chipset Intel Z77 Express, which we set overclocking processors Core i5 with designs Ivy Bridge and Sandy Bridge. But the main role in the study, depending on the setting of performance memory kit got high-speed memory standard, DDR3-2600, granted to us by G.Skill.
In general, the tests were used the following hardware and software components:
- Processors:
- Intel Core i5-2550K, overclocked to 4.5 GHz (Sandy Bridge, 4 core, 6 MB L3);
- Intel Core i5-3570K, overclocked to 4.5 GHz (Ivy Bridge, 4 core, 6 MB L3).
- CPU cooler: NZXT Havik 140;
- Motherboard: ASUS P8Z77-V Deluxe (LGA1155, Intel Z77 Express);
- Memory: 2×4 GB, DDR3-2600 SDRAM, 10-12-12-31 (G.Skill TridentX F3-2600C10D-8GTXD).
- Graphics Card: NVIDIA GeForce GTX 680 (2 Gbayt/256 bit GDDR5, 1006/6008 MHz).
- Hard Disk: Intel SSD 520 240 GB (SSDSC2CW240A3K5).
- PSU: Tagan TG880-U33II (880 W).
- Operating System: Microsoft Windows 7 SP1 Ultimate x64.
- Drivers:
- Intel Chipset Driver 9.3.0.1019;
- Intel Graphics Media Accelerator Driver 15.26.12.64.2761;
- Intel Management Engine Driver 8.0.0.1399;
- Intel Rapid Storage Technology 11.1.0.1006;
- NVIDIA GeForce 301.42 Driver.
Note that in this test, we used only overclocked to 4.5 GHz processors. The fact that the increase of clock frequency provides a more pronounced picture of the dependence of performance on the parameters of the memory subsystem.
Features of the memory controller Ivy Bridge
The evolution used in the Intel Memory Controller has been a long and multistage. However, it seems, so far, engineers have come to the optimal scheme in terms of architecture of the memory controller Ivy Bridge and Sandy Bridge are virtually indistinguishable. As a basis for the smooth functioning of the new controller was introduced back in the ring Sandy Bridge design on-chip bus. All computing and graphics cores, it has provided equal access and the fastest route to a third-level cache, as well as to the actual memory. As a result, the peak data rate increased significantly, and the Core for LGA1155-systems are significantly ahead in the practical tests the memory of all its competitors.
Quite surprisingly, in the design of Ivy Bridge principles of interaction between the processor and memory controller remained unchanged.Moreover, the engineers did not have to climb in the interior of the controller and all the changes are cosmetic in nature, especially as Ivy Bridge processors are fully compatible with the existing platform LGA1155. So again we are dealing with the familiar two-channel controller designed to work with DDR3 SDRAM. Compared with Sandy Bridge grew only officially supported frequencies – in the characteristics of performance now means Ivy Bridge to DDR3-1333 and DDR3-1600, a new version of the XMP technology has been numbered 1.3. However, the first thing the second – the usual “paper” improvements. In practice, they absolutely do not mean anything, its essence lies in the mainstream of formal specifications.
Just as before, the memory controller in the Ivy Bridge can operate in symmetric mode, when the volume, frequency and delay modules in both memory channels are the same, and in compatibility mode, called the Intel Flex Memory Technology. The technology Flex Memory is that, based on the characteristics of the modules, the entire memory array is divided into two parts: one to which we can apply the symmetric mode of access, and another – which should work in single-channel asymmetric mode. As a result, LGA1155-system can be completed in pairs differing sets of DIMM modules in different channels, but the performance drop is not catastrophic.
Each channel memory controller Ivy Bridge is able to work with one or two single-sided or double-sided modules, DDR3 SDRAM, and this means that the maximum supported by the latest LGA1155 system memory is 32 GB. For those who need large amounts of memory, Intel is ready to offer more high-end platform – LGA2011.
All of the above in this section so far, in fact, a simple retelling of typical properties of Intel’s memory controller and equally valid not only for the Ivy Bridge, but for Sandy Bridge. However, with the introduction of a fresh processor design in the memory controller and the original features have appeared due to the transition to an improved scheme for its frequency.
Fundamental changes in two. First, the variable factor is responsible for obtaining the resulting memory frequency, was more freedom. While Sandy Bridge for maximum achievable frequency mode memory performed DDR3-2400 processors Ivy Bridge opened and faster modes – now the DDR3 SDRAM can be clocked at frequencies up to 3200 MHz. Second, in the scheme of setting the memory frequency has an additional variable coefficient allowing for the usual 266-megahertz discrete add another option changes the frequency of DDR3 SDRAM – 200 MHz increments.
As a result, the cycling of memory has become much more flexible. Of course, changing the base frequency clock (BCLK) in LGA1155-system does not work, yet the memory frequency, you can choose from a variety of preset values. So, for example, is now a list of available modes to choose DDR3 SDRAM in LGA1155 processor platform family of Ivy Bridge.
It should only be kept in mind that this screenshot is on the motherboard with a chipset Z77, which was installed CPU Core i5-3570K. The boards also, which are based on chipsets H-series, this flexibility in configuring the frequency of DDR3 SDRAM does not provide. They are the choice will be limited only by regular DDR3-1333 and DDR3-1600. Another limitation concerns processors Ivy Bridge, not related to overclocking K-series. They are free to increase the memory frequency to 2400 MHz only, as a high-speed mode – not work.
But in systems with processors Core i5-3570K and Core i7-3770K overclocking DDR3 SDRAM offers the full expanse. As shown by our experiments, high-speed modes are fully functional, and for using them do not even need any kind of tedious selection tricks secondary voltages. For example, in our test system for the stable functioning of the Ivy Bridge with DDR3-2667 SDRAM was quite a small (only 50 mV), increasing the voltage on the memory controller.
The simplicity of achieving higher memory frequencies and shows herself Intel. Stability in this case can be improved by changing only two voltages. Voltage V DDQ , which is fed directly to the modules. But he was, as before, is not recommended to raise above 1.65 in order to prevent damage or degradation of the processor. And the voltage V CCSA , is responsible for feeding the system agent and the memory controller. This value is the nominal value of 0.925 V, and a slight excess improves the stability of modes with high frequency DDR3.
As a result, we can conclude that innovations are implemented in the controller memory processors Ivy Bridge, aimed primarily at overclockers who prefer to use the DDR3 SDRAM high in abnormal conditions. As for the memory controller in normal mode, Intel does not promise any change in comparison with Sandy Bridge.
However, to make absolutely clear, we decided to compare the practical performance speed controller Sandy Bridge and Ivy Bridge and when they work with “ordinary” two-channel DDR3-1600 with timings of typical 9-9-9-27-1N. Testing was conducted in the same LGA1155-platform, which replaced the processors only. To the performance of their controllers are not influenced by various clock speeds, and 22-nm and 32 nm CPU was overclocked to the same frequency of 4.5 GHz. All energy-saving technologies, as well as technology Turbo Boost, disconnected.Given the possibility of forming a memory frequency of 1600 MHz processors Ivy Bridge in two ways (1600 MHz DDR = 100 MHz x 1,33 x 6 or 1600 MHz DDR = 100 MHz x 1,00 x 8), the tests were subjected to both.
The results of the measurement bandwidth and latency of the memory subsystem are listed in the table below:
|
Sandy Bridge 4.5 GHz DDR3-1600 9-9-9-27-1N |
Ivy Bridge 4.5 GHz DDR3-1600 9-9-9-27-1N (1.33x Memory Ratio) |
Ivy Bridge 4.5 GHz DDR3-1600 9-9-9-27-1N (1.00x Memory Ratio) |
|
|---|---|---|---|
| Aida64 2.50.2018 Cache & Memory Benchmark | |||
| Read, MB / s | 19,480 | 19,808 | 19,801 |
| Write, MB / s | 22,165 | 22,096 | 22,105 |
| Copy, MB / s | 21,451 | 21,665 | 21,728 |
| Latency, ns | 45.4 | 41.3 | 41.3 |
| Sandra 2012.SP4c | |||
| Bandwidth, GB / s | 21.65 | 21.78 | 21.76 |
| Latency, ns | 19.5 | 19.9 | 19.9 |
| STREAM 5.8 | |||
| Bandwidth, 1 Thread, MB / s | 21,497 | 21,480 | 21,393 |
| Bandwidth, 4 Threads, MB / s | 22,295 | 22,394 | 22,490 |
The first important conclusion to be drawn from the numbers obtained, it is – almost complete independence of the performance of the memory controller on your Ivy Bridge “extra” factor. Cycling of memory in increments of 200 or 266 MHz gives very similar results. Practical indicators of performance in different variants of the formation of memory frequencies differ by no more than a few tenths of a percent, which can be attributed simply to measurement error.
As for the relative performance of the memory controllers Sandy Bridge and Ivy Bridge, it seems, is not quite identical. While the capacity of the memory subsystem as a whole is not much different, the differences in the practical latency can reach a few percent. At the same time say for sure which of the controller is faster – absolutely impossible. It all depends on the nuances of memory, so that in general we can say that in terms of operations with the RAM or the owners of Sandy Bridge, Ivy Bridge or the holders of any specific benefits do not receive.
Frequency vs. Timing
Whenever it comes to choosing the optimal memory, sooner or later the question arises, why should strive for in the first place: to increase the frequency of the memory or to reduce delays. But this time, we avoid the detailed test modules DDR3 SDRAM, differing only timings. The fact that with each new platform impact of delays on the overall performance decreased, and so far it is, perhaps, has already passed the critical point. Of course, the dependence of productivity on the timings are not fully come to nothing, but compared to the effect that has on the performance of the system change in frequency of DDR3 SDRAM, she was unobtrusive.
There are two main reasons. First, with increasing frequency of the memory of her minimum latency, in any case increases, and against this background that the relative value increments varying delays is becoming less and less noticeable. It is one thing – increase timing a couple of cycles with three or four (as in the case of DDR2 SDRAM), and the other – with seven or eight (in the case of DDR3 SDRAM). In the first case, the latency increased by 50-70 percent, while the second – only 25-30 percent. Accordingly, the difference between the different variants of modern memory timings have been nowhere near as important as before.
The second reason – it is the general scheme of improvement of processor memory. There is growing not only the number of levels of caching data, but the amount used by modern processors, cache memory. All of this masks the real latency of memory, shifting the emphasis on its capacity.
Actually, no need to race for low latency for high-frequency DDR3 SDRAM long prosekli and manufacturers of overclocker memory kits. Offers low-latency disappeared from the market, and is now on store shelves is difficult to find a DDR3 SDRAM modules with a latency of less than 9 cycles. The number of proposals with ultra-high frequencies and long delays while steadily increasing.
However, we do not want to leave unfounded assertions about the insignificance of the influence of timing on the memory subsystem performance in a modern platform, built on processors Ivy Bridge. Therefore, we carried out and the practical test by which to compare the actual performance of identical systems with overclocked to 4.5 GHz Core i5-3570K, equipped DDR3-1600 and DDR3-1867 SDRAM with different delays.
The graphs are the most vivid illustration of the above. The increase in memory frequency at 266 MHz is more effective than the reduction of delays for 3-4 cycles. And in terms of actual latency even DDR3-1867 with timings of 9-9-9-27 is better than DDR3-1600 7-7-7-21 latencies to aggressive. If you judge the speed of the memory subsystem, based on indicators of the actual capacity, the DDR3-1600 can not be compared with a slightly higher frequency option at all under any circumstances.
All this means that the memory latency in modern systems are quite insignificant parameter. Therefore, the choice of DDR3 SDRAM Processor Ivy Bridge in the first place you should pay attention to the frequency of its operation, and low CAS Latency, and similar values have practically no effect on the actual performance. Similarly, should come and set up and overclocking – you must first fight for the increase in the frequency of the DDR3 SDRAM, and only then, if you wish to engage in minimizing delays.
The dependence of the frequency of memory
In the meantime, we come to the culmination of our material: to clarify how memory settings can affect the overall performance of the platform in real-world problems. In this case, given in the previous section made the findings, we decided to abandon the test problems commonly used in the memory subsystems with different timings alone, and made the focus to the much more explicit dependence on the frequency of performance DDR3 SDRAM. For comparison, we selected common memory configuration with a frequency of 1333 MHz to 2667, which established the most common delays. Specifically, this means that in comparison attended the following options for DDR3 SDRAM:
- DDR3-1333 9-9-9-27-1N;
- DDR3-1600 9-9-9-27-1N;
- DDR3-1867 9-9-9-27-1N;
- DDR3-2133 11-11-11-33-1N;
- DDR3-2400 11-11-11-33-1N;
- DDR3-2667 11-13-13-35-1N.
The rest of the test system, based on overclocked to 4.5 GHz quad-core Core i5-3570K with a micro-Ivy Bridge, absolutely nothing has changed.
Increasing the frequency of the DDR3 SDRAM is to be expected can raise the performance of practical bandwidth and reduce latency in practice. At the same time quite interesting that the largest increase in memory speed is observed with increasing frequency up to DDR3 SDRAM 2133 MHz in the future is the impact of high frequency slightly weakened. But the most significant leap in results occurs at the first stage in the transition to 1600-MHz memory, and it’s as if hinting that DDR3-1333 to the present conditions it is time to be attributed to the legacy proposals. In general, doubling the memory frequency to 1333 MHz up to 2666 results in no more than 50 percent and increase the capacity of actually observed. At about the same scale changes and the practical latency.
However, the memory benchmark from Aida64 is characterized in that the character is single-threaded, because of what it reveals the potential of modern memory controllers are not in full. Therefore, we used an additional benchmark, and Stream is single-threaded and chetyrehpotochnom (number of cores) mode.
Indeed, here the dependence of bandwidth on the frequency of memory are more pronounced than in the benchmark Aida64. Overclocking memory to 2133 MHz to 2400 MHz gives a very noticeable effect, but the next 266-MHz frequency step does not look just as useful.Nevertheless, the resulting increase in the practical processing speed when moving from DDR3-1333 to DDR3-2400 or DDR3-2666 up to 64 percent.
Synthetic tests depict the “ideal” picture, but give no idea of how it will depend on the memory clock speed of the system in the usual programs. So then we went to the tests in complex benchmarks and real applications.
In the synthetic memory tests all looked very nice, but Futuremark benchmarks paint a picture of the mundane. The frequency of memory continues to have an impact on system performance, but the difference in results is not too impressive. Increasing the memory frequency at 266 MHz raises integral indicators PCMark and 3DMark 11 7 less than one percent, a significant sensitivity to the same parameters of the memory shows only the physical subtest graphics benchmark. It due to overclocking DDR3 SDRAM can get up to 14 percent additional performance.
Most of the frequency dependence of the performance of the memory is observed in the archiver WinRAR. In other applications as fast DDR3 SDRAM is able to provide an increase to speed within a few percent.
However, the situation evolving in gaming applications, from the overall picture is somewhat different after all. In gaming systems, the speed of the memory subsystem has a greater than usual importance. And the choice of platform, built on a processor Ivy Bridge, DDR3-memory with high bandwidth can provide additional 5-10 percent performance in games, to get that is not always possible, even installing a faster processor models.
The frequency of memory and performance of integrated graphics
Built-in video core processors today operate for their own use system memory along with the computational resources of CPU. Therefore, their graphics performance also depend on the speed of the platform, installed in DDR3 SDRAM. Moreover, in the 3D-rendering is a very intensive exchange of information with the texture memory, so the impact of its speed must be at least no less prominent than in the case of traditional computing performance. That is why we decided to pay special attention to the study of performance graphics core Intel HD Graphics 4000 system memory when working on a different frequency.
Curiously, 3DMark 11 does not detect a significant index of subordination graphics performance memory bandwidth. The difference between the best (with DDR3-2666 SDRAM) and worst (with DDR3-1333) the result is only 2.5 percent, even less than the difference in computational performance in many non-graphical applications. Apparently, the micro-Ivy Bridge, in which the graphics core, not only has its own cache memory, but can use the L3-cache processor, eliminates the effect of memory bandwidth at the speed of the GPU HD Graphics 4000/2500.
However, in real games can be seen a slightly different picture.
The speed of the games on the built-in graphics processors Ivy Bridge on the frequency of memory still depends. The influence of this parameter is very important: every additional 266 MHz translate into increasing the number of frames per second by a few percent, as a whole is able to overclock the memory increase, sometimes reaching the 25 percent value. However, this gain is characterized by an increase in speed graphics core is only partly, because, as we saw earlier, higher memory performance leads to an increase in gaming performance and the use of an external video card. However, if you intend to actively use the built-in processor Ivy Bridge graphics core, to neglect the possibility of increasing its efficiency by installing faster DDR3 SDRAM is clearly not to be.
