Results of testing memory modules at a system bus frequency of 266 MHz
Until now, unit testing DDR memory 2-533 (and even DDR2-667) were carried out on motherboards, based on Intel 915/925 series chipsets operating at a system bus frequency of 200 MHz (800 MHz Quad-Pumped Bus). At the same time, there was a quite obvious containment of the real potential of this type memory, due to the fact that the peak throughput The 200-MHz bus system is only 6.4 GB/s, while the native theoretical bandwidth of DDR2-533 in dual-channel mode is 8.53 GB/s. Of course, such a limitation did not arise when testing this type of memory in single-channel mode, in which it revealed its maximum potential, as we wrote about earlier.
Our testing laboratory received engineering samples of the new Intel Pentium 4 Extreme Edition 3.46 GHz processor and the Intel D925XECV2 motherboard, based on the recently announced Intel 925XE chipset, which is essentially a 266 MHz variant of the Intel 925X. A separate article will be devoted to the consideration of this processor, motherboard and chipset as such. The purpose of this small study is to show what should be expected from the transition from a 200 MHz to a 266 MHz processor bus in terms of performance of a memory subsystem with DDR2-533 modules operating in dual-channel mode. To solve this problem, we, as usual, will measure the low-level characteristics of the memory subsystem (bandwidth and latency) using the RightMark Memory Analyzer test package.
Test bench configuration
Test bench No. 1
- CPU: Intel Pentium 4 Extreme Edition 3.4 GHz (Gallatin core)
- Chipset: Intel 925X, FSB frequency 200 MHz
- Motherboard: Intel D925XCV, BIOS version 1259 from 08/19/2004
Test bench No. 2
- Processor: Intel Pentium 4 Extreme Edition 3.46 GHz (Gallatin core)
- Chipset: Intel 925XE, FSB frequency 266 MHz
- Motherboard: Intel D925XECV2, BIOS version 1012 from 09/23/2004
- Memory: 2x256 MB Samsung DDR2-533, timings 4-4-4-11
- Video: Leadtek PX350 TDH, NVIDIA PCX5900
- HDD: WD Raptor WD360, SATA, 10000 rpm, 36Gb
- Drivers: NVIDIA Forceware 62.01, Intel Chipset Utility 6.0.1.1002, DirectX 9.0c
Test results
The testing involved two similar platforms based on Intel processors Pentium 4 Extreme Edition (512 KB L2 cache, 2 MB L3 cache, Gallatin core), Intel motherboards based on 925X and 925XE chipsets with two 256 MB Samsung DDR2-533 memory modules operating in dual-channel mode. In the settings of the memory subsystem, the standard timing scheme 4-4-4-11 was used, prescribed in SPD chip modules.
Real Memory Bandwidth
On the first platform (with a FSB frequency of 200 MHz), the average real read bandwidth is 4065 MB/s, i.e. approximately 63.5% of the maximum theoretical memory bandwidth of DDR2-533, limited theoretical processor bus bandwidth of 6.4 GB/s. Moving to a 266 MHz processor bus, removing this limitation(since its theoretical PS is already 8.53 GB/s, which is equal to the theoretical memory PS), increases it to ~4500 MB/s. The absolute increase is small - only 10.6%, and the relative one - completely negative (52.7% of the theoretical memory bandwidth versus 63.5% on the first test bench). An even less significant absolute increase is observed in the average real bandwidth per write (1895 MB/s versus 1780, i.e. only 6.4%). However, these values are “average”, they only indirectly relate to the real characteristics of memory bandwidth, because they are limited by many other factors, primarily the processor architecture itself (in particular, in tests of average real memory bandwidth on recording, the negative impact of the feature is great processor cache operation for writing).
Maximum real memory bandwidth DDR2-533,
Moving on to the maximum values (the corresponding curves are presented on the graph), we note that in reality they are not particularly “maximum”, because they are also really limited by the processor architecture (read bandwidth efficiency of the Software Prefetch algorithm, write bandwidth efficiency of direct access to memory, bypassing the processor cache). Maximum real read bandwidth on the first platform 5388 MB/s (84.2% of the theoretical maximum 6.4 GB/s). The 266-MHz bus increases the value of this parameter to 6366 MB/s, which is 18.2% higher in absolute value, but again lower in relative units (74.6% of the theoretical maximum of 8.53 GB/s).
Unlocking the real potential of DDR2-533 memory bandwidth, oddly enough, can be seen only by the values of the maximum real memory bandwidth per write, which, according to our numerous studies, is strictly limited at the level of 2/3 of the theoretical memory bandwidth of the processor bus. This is what is observed in both cases: 4267 MB/s (66.7% of the theoretical maximum) on the first platform, 5674 MB/s (66.5% of the theoretical maximum) on the second. The increase is 32.9%, which is very close to the expected (in an ideal case) 33.3%. So, in our first series of tests, even if only in one case, and even then “indirectly,” we achieved victory for the 266-MHz processor bus of the i925XE chipset over the 200-MHz version of the i925X when using DDR2-533 memory in dual-channel mode.
Memory latency
The methodology for measuring latency, as applied to processors of the Pentium 4 family, was developed in detail, justified and described earlier. Therefore, we will dwell on it only briefly: the latency test uses pseudorandom(as well as completely random) mode of traversing a relatively large block of memory ( 16 MB) in increments of 128 bytes(“effective” size of an L2/L3 cache line, associated with hardware prefetching of an adjacent line from memory to the cache in all bypass modes).
DDR2-533 memory latency (pseudorandom and random bypass),
system bus frequency 266 MHz
In fact, latency values obtained from a truly random walk through the allocated memory block are not very meaningful due to the fact that a large component of this parameter is the D-TLB miss value. Which, by the way, is clearly visible in the form of the difference between the pseudo-random and random access latency curves shown on the graph. Therefore, hereinafter, by “memory latency” we will understand the latency of the pseudo-random traversal of the chain.
* without tire unloading
** block size 16 MB
The average memory latency (obtained without unloading the bus by inserting “empty” operations) on the first platform, in which the memory operates in asynchronous mode, is 81.6 ns. The spread of latency values under conditions of gradual bus unloading is from 79.4 to 119.9 ns. Switching memory to synchronous mode (second platform) has a very positive effect on latency - in all cases it decreases by 9-10 ns. The same picture, by the way, is observed in the case of random access latency (the same spread of values and their decrease by 9-10 ns in synchronous mode). Finally, it should be noted that the latency values themselves are very high, which is due to the type of processor core used (Gallatin, which is a variant of the Northwood core with 2 MB L3 cache), which does not have the most efficient hardware data prefetch algorithm (Hardware Prefetch), and also not the most efficient implementation of BIU (cache memory bus).
Results
A few months ago we wrote that using DDR2-533 memory will really pay off only with the advent of chipsets that support a 266 MHz processor bus. According to the results of our testing today, this is true... but only partly, so it’s time to make a small clarification. So, not only the frequency (directly related to the bandwidth) of the processor bus is important; the efficiency of the implementation of memory logic on the processor side is also important. As is known, in particular, from our tests, Northwood (Gallatin) processor cores do not have such high efficiency of Hardware and Software Prefetch algorithms and BIU functioning, which is achieved with the Prescott core. In this connection, full disclosure of the real potential of DDR2-533 in dual-channel mode will become possible only with the release Pentium processors 4 on the Prescott core, supporting a 266 MHz processor bus frequency (1066 MHz Quad-Pumped Bus). We will definitely return to this moment when the first samples of such processors become available to us.
GBT GeForce PCX 5900 Let me remind you that the i925X and i915P chipsets support the new kind memory - DDR II, which surpasses DDR I in bandwidth. However, due to higher timings, overall system performance remains at the same level as when using DDR I memory.So, both chipsets (i925X and i915P) support DDR2-400 and DDR2-533 memory. This means that with a system bus (FSB) frequency of 200 MHz, in the first case the memory operates at a frequency of 100 MHz, in the second - 133 MHz (and due to the features of the DDR2 architecture, this value is quadrupled and we get 400 and 533 MHz, respectively).
We also note that Asus company announced support for DDR2-600 memory in its motherboards. In particular, P5AD2 Premium (i925X) and P5GD2 (i915P) have a corresponding parameter in the memory frequency setting section. In other words, the company's engineers have opened access to an unofficial frequency divider = 34 (MEMCPU).
But in practice we were unable to run the system with this divider.
The first set of memory modules that appeared in our laboratory was Kingmax DDR2-533. 
To enlarge, click (~170Kb).
Please note that Kingmax uses its own branding chips to produce these modules.
According to SPD, these modules have the following timings: 3-3-9-3 for a frequency of 100 MHz, and 4-4-12-4 for a frequency of 133 MHz. At standard frequencies, the modules performed all tests flawlessly. And now we will try to find out what the potential of this memory is when overclocked. For this purpose we have assembled a stand with Asus board P5GD2 on the i915P chipset and Pentium4 3.2 GHz processor. The maximum possible FSB frequency at which the system worked stably was 240 MHz. This result is slightly lower than Asus board P5AD2 (i925X; max. FSB=250MHz), but unfortunately the latest BIOS for this board (v 1004) does not allow you to manually change memory timings.
As a result, when overclocking with a multiplier of 12 (menu item "DDR2-400") the maximum memory frequency is 120 MHz (DDR2-480), and when using a multiplier of 23 (menu item "DDR2-533") the frequency is 160 MHz (DDR2-640) .
So, the results:
As we can see, Kingmax DDR2-533 memory worked great at minimum timings, up to the highest possible frequency (specifically for this system). Naturally, when the timings were increased, the results did not change. In addition, there is a significant margin for increasing frequency. For example, on the Asus P5AD2 board, this memory operated at a frequency of 166 (DDR2-667) MHz (FSB = 250 MHz; SPD timings).
How good is this result? Maybe all other DDR2 modules will demonstrate the same performance. To clarify this issue, we tested a couple more DDR modules 2-533 of an unknown manufacturer (assembled on Samsung chips).
Samsung chips 
As it turned out, these noname modules showed significantly worse results. In particular, at minimum timings, the maximum possible memory frequency was 120 MHz (maximum when using a divider of 12). And when using a multiplier of 23 and FSB frequency = 200 MHz (i.e. memory frequency = 133 MHz DDR2-533), the system could only start, but could not load Windows. The situation was somewhat improved by increasing the Vmem voltage to 2.0V - Windows booted and the tests were successful. But increasing the memory frequency by literally 1 MHz led to malfunctions.
Actually, we have determined the potential of both sets of memory. But how to use this information in practice? Questions immediately arise: “Which mode is the most productive: DDR2-400 or DDR2-533?”, “How much does performance change when timings change?”
To answer these questions, we conducted another series of tests. The results are on the next page.
| Content: | ||
| Page 1 - Memory Kingmax DDR2-533 Page 2 - Optimizing DDR2 memory on i925Xi915P chipsets |
Optimizing DDR2 memory on i925Xi915P chipsets
So, the first question is: “Which mode is the most productive: DDR2-400 or DDR2-533?” To do this, we set the standard frequency FSB = 200 MHz, timings according to SPD.
| PCMark 2002 (cpumem) | Quake3 (fastest ; fps) | Winrar (KBytes) | ScienceMark (MBytes) | |
| Asus P5AD2; DDR2-400 (SPD) | 681610558 | 456.1 | 398 | 4089.81 |
| Asus P5AD2; DDR2-533 (SPD) | 680410472 | 456 | 398 | 4086.32 |
| Asus P5GD2; DDR2-400 (SPD) | 682510204 | 468.8 | 398 | 4052.54 |
| Asus P5GD2; DDR2-533 (SPD) | 680810264 | 469.8 | 388 | 4027.34 |
The first two lines of results were obtained on a system with an Asus P5AD2 board on the i925X chipset and BIOS 1004.
The second two lines are on the P5GD2 board on the i915P chipset with BIOS version 1001.
As can be seen from the table, the performance of both modes is approximately the same. It turns out that it is not at all necessary to force the memory to work in DDR2-533 mode. Surely we will get higher performance in DDR2-400 mode and lower timings.
The fact is that the bandwidth of DDR2-400 fully satisfies the needs of the Pentium4 processor running on a 200 (800QPB) MHz bus. The processor bus bandwidth is 6.4 GB, and prop. DDR2-400 memory capacity is 6.4GB. And when using DDR2-533, the memory bandwidth increases to 8.5 GB, but the processor bus pumping speed does not change. As a result, the user does not receive any increase in speed.
For clarity, I will give the following example: if we need to fill a 2-liter container from a kitchen tap (naturally, completely open), then we will spend the same time if we use a plastic bottle with a neck corresponding to the diameter of the tap (DDR2-400), or a mug with a large neck (DDR2-533), that is, performance depends on the diameter of the water tap (or processor bus.
Now let's see how performance depends on memory timings (FSB=200MHz; memory frequency divider=12 or "DDR2-400"):
| PCMark 2002 (cpumem) | Quake3 (fastest ; fps) | Winrar (KBytes) | ScienceMark (MBytes) | |
| 3-3-8-3 | 680010242 | 449,4 | 390 | 4025,2 |
| 3-3-9-3 | 681810225 | 470,3 | 389 | 4106,77 |
| 3-4-8-4 | 680010060 | 459 | 360 | 4081 |
| 3-4-12-4 | 680910067 | 454,4 | 351 | 4079,48 |
| 4-3-8-3 | 680010079 | 446 | 378 | 4023,68 |
| 4-4-8-4 | 68059893 | 452,9 | 350 | 3987,71 |
| 4-4-12-4 | 67829896 | 430 | 343 | 3970 |
| 5-5-15-5 | 67769502 | 413 | 306 | 3829,31 |
Standard timings are shown in bold.
In general, the picture is clear: the lower the memory timings, the higher the system performance. But there are also small subtleties. For example, if we reduce the timings from 3-3-9-3 to 3-3-8-3, then the performance will decrease (and quite noticeably). This indicates that with non-standard timings, some disturbances occur in the synchronization of data signal transmission, which leads to slower operation. As a result, setting timings at random can seriously slow down your computer.
What do we get in the end?
1) For the average user, you can absolutely safely set the memory modules to DDR2-400 mode (i.e., memory frequency divider = 12), as well as SPD timings. As a result, the system will perform at least no worse than with DDR2-533 memory. And in most cases, the system will work faster due to lower timings.
2) For experienced users overclocking their system, DDR2-400 mode is also recommended. Due to lower timings, we get higher performance. And due to the higher memory frequency divider, we are at high frequency FSB has a relatively low memory frequency that operates at low timings.
From this point of view, Kingmax DDR-533 memory looks very attractive. The most important thing is that it remains operational at frequencies of 160 MHz (DDR2-640) and higher, with the lowest possible timings. In this case, the memory modules operate at standard voltage, which means an increase in overclocking potential as the voltage Vmem increases.
From experience using DDR I modules, we know that memory compatibility with various motherboard models plays a big role. It happens that the same memory shows completely different results on boards various manufacturers. There are also often cases where memory performance depends on different BIOS versions for the same motherboard.
It is clear that this situation has little concern for ordinary users - they receive a ready-made computer from the company, and all the headaches of selecting memory fall on the assemblers, from this very company. But overclockers will have to sweat when choosing the right combination" motherboard+ RAM".
All questions, comments and suggestions can and should be asked at.
Greetings, my dear readers and visitors! 🙂
Today I would like to consider an issue that, judging by the letters I receive, worries a considerable number of users. Namely, the question about the speed standards of DDR2 RAM: are the differences between them significant? what will be optimal? and etc…
So, today DDR2 reigns on the market; there are several main speed standards for this memory. DDR2-533, DDR2-667, DDR2-800, DDR2-1066. There are others, but these are so-called overclocking standards, the use of which only makes sense when the system is overclocked to extremes.
So, DDR2-533. The lowest speed standard. Why 533? 533 is the module operating frequency in MHz. Today there is not much memory of this standard (available for free for desktop PCs) - it is leaving the market. Why? Because today we can see probably the lowest prices for RAM DDR2. There are many manufacturers, many models, a lot of memory, etc. In general, there are many reasons, but that’s not what I’m talking about now... DDR2-533 was previously relevant in light of its extreme cheapness, but today this is a little different - all memory is cheap. DDR2-533 was used mainly in office and low-cost home PCs. Today, memory of this standard is used in some laptops.
DDR2-667- average. In a good way. In 2007, memory of this standard was the most common in home universal desktop computers. It already operates at a frequency of 667 MHz. This memory still sells very well today and is usually installed in universal desktop PCs. It is also used in laptops.
DDR2-800- "Older brother". Memory of this standard is very popular this year, 2008, for installation in universal, home and simply high-performance PCs. This was mainly due to a significant reduction in the price of memory of this standard. This memory operates at a frequency of 800 MHz. It can also be installed on laptops.
DDR2-1066- high-performance memory. Not much different from the 800. The only difference is better overclocking potential. Accordingly, this memory is very popular among overclockers who install it in their powerful PCs. When overclocked, it provides even greater performance compared to DDR2-800.
So, brief description given. What are the differences? And the differences, to be honest, are not that strong...
If the memory operates in normal mode, the PC is not overclocked by the user (or overclocked, but slightly), then as a rule there is only a significant difference between DDR2-533 and DDR2-1066 modules. And the difference between DDR2-667, for example, and DDR2-800, one might say, is not at all noticeable in everyday work. Why? This is related to the organization of the memory subsystem and the principles of its operation (I won’t burden you with this, I’ll just mention it 😉). The same can be said about DDR2-533 vs DDR2-667.
As I already mentioned, memory is very cheap these days. Previously (in my opinion, even at the beginning of 2007), prices for memory of different standards differed quite significantly, and you could be faced with a choice - which would be optimal. Today the advice is this: since the price of memory is low, it is better to take, as they say, “to the maximum.” Extra memory performance won't hurt (hmm, if you can even call it extra...). Therefore, in most cases today optimal solution for installation in desktop PCs (home, medium/high-performance) is DDR2-800. On average, installing a 2 GB memory kit (800) will cost today on average 1500 rubles.
The only correction: if you are an overclocker and plan to overclock your computer, then it is better to buy DDR2-1066 memory, it will provide the necessary level of overclocking potential.
P.S. Of course, it would be stupid to choose RAM only based on the speed standard! There are still many nuances in choosing memory, such as timings, etc... Therefore, when choosing memory, you must definitely take all this into account (all the points and nuances) and seek help from a specialist.
P.P.S. DDR3 is gradually entering the market. What is this? Is it profitable to use this memory?... I will soon try to answer all these questions on the pages of this site.
Well, as they say, once such a p... uh... sorting out began. 🙂 First, the total of the 1st episode; objects (folders, shortcuts, etc.) seemed to be nailed down and were not moved by any click, “insert” stopped working context menu(always inactive), these same errors were not clicked in the error logs to see the description when logging into Accounts an empty window without selecting anything, in the task manager the absence of a loved one in the users tab and generally loss of administrator rights, partial or complete hs (message when trying to launch an application on drive D), processes in the task manager instead of +50 left 30+, periodic reboots with a blue screen (quickly flashing, you don’t have time to look at what’s written there), later we managed to figure out the error code
Error code 10000050, parameter1 8f640cec, parameter2 00000001, parameter3 805b641a, parameter4 00000000.
Error code 10000050, parameter1 c399ff20, parameter2 00000000, parameter3 bf80dd9b, parameter4 00000000.
something like this, when I try to scan for viruses, there is also a reboot (in fact, I tried to fight them for 3 days), messages about a broken file system in C, and so on and so forth. The main problem was to remove texts with passwords/logins. I was already mentally ready to rewrite it manually, but remembering about the Windows disk, I successfully used the file transfer wizard. (fine-soft ones are not as bad as they really are =))) I don’t remember how it all started, but definitely after that As soon as I started manipulating the memory, I still remember there was something freezing, scandisk and off we go. I tried to restore the system - again an error and a reboot. (now in the pad I write Ctrl+S after each sentence, because the reptile regularly reboots:(). Everything described was running with home edishin, the second XP (cut off from game edit) also almost didn’t start at all, complaining about the broken C. C safe mode nothing good came of it either. Having pushed around, I pulled up heavy artillery and Acronis True Image Home 11.0 restored logical C sector by sector. Everything seemed to work normally (although right now there is such confusion in my head that I can’t guarantee anything :)) And the second axis started working. I exchanged the memory (goodram) I think maybe the bracket was buggy. I inserted it, everything seemed to be fine in PC Wizard 2008, I even tested it, it showed something like my old 4200. Oh well, I connected to DSL and let’s download new things. The image of Acronis was already in October 2008, albeit with almost all the necessary programs. Well, here I am sitting here, stuffing an iron friend... and bam. Again the old song. There hasn't been a reboot... mother... for a long time. Similar codes, application error log is already corrupted. Something was freezing (again, beyond my memory:), Scandisk was checking something there. True, this time there was no folder on the disk where there is 000 at the end.
So I'm back again after the reboot. :) Some crap wanted to go to the Internet (it's disabled), I banned it in Komodo. Then I went into it to see in more detail what it was, clicked in the log... an error window and a reboot. After the error message savedump.exe and now there is no record of this event. Somehow I don’t even know what to think. Maybe it really is some kind of virus. Maybe some idiot (I can’t hold back anymore) registered in the MBR? Well, Acronis is registered there (recovery at boot). True, launching it with the choice F11 (recovery) 2-3 times yesterday, and even now displays MBR error 2. Maybe there’s something wrong here? In short, I have no strength. I lay it out and go to bed. Tomorrow (today) I’ll restore it again with aronise and see how it develops with the old memory. PS By the way, the day before I fitted the mouse with a double-click button... Maybe there’s something here? =)))))) ZYY I’m stuck, I can’t tear myself away. Overloaded again. And again I got into some kind of small-soft synchronizer. Something like this. ZYYY I couldn’t go into reboot with firelis, I spat and installed my RAM. It seems to last for some minutes. :) That memory was so hot...even though it’s a laptop.
Loading...
