Choosing Optimal Memory to Match Intel Pentium 4 Processor

Choosing proper memory for your Pentium 4 based system is a pretty complicated task. We did our best to answer the most frequently asked questions regarding this matter. How do memory timings affect the overall system performance? Would it be better to use high-speed overclocker memory for CPU overclocking? What is more important for Pentium 4 systems: high memory working frequency or low latency? Find all answers in our article!

by Ilya Gavrichenkov
09/15/2003 | 11:46 PM

Intel’s transition of the Pentium 4 processor family to the 800MHz system bus was an important event not for the CPU market only, but for the memory market as well. Pentium 4 CPUs with 800MHz FSB required new chipsets.

 

We have them now – the i875 and i865 chipset families, which also allow using dual-channel DDR400 SDRAM for the memory subsystem. Moreover, the use of DDR400 SDRAM with these processors is not only possible, but even necessary to reach the highest performance. It’s not only because DDR400 SDRAM and the 800MHz bus can work synchronously, thus eliminating the latencies that occur otherwise.

The main advantage of this combination is that the bandwidths of the memory and system buses match each other. The 64-bit Quad Pumped Bus of the modern Pentium 4 CPU working at 800MHz provides a bandwidth of 6.4GB/s. This is the exact match of the bandwidth of two DDR400 SDRAM channels. This is the main reason for us (and for Intel) to say that the dual-channel DDR400 if the optimal memory subsystem for the top-end Pentium 4 processor today.

It should also be mentioned that Intel spends much effort to promote the 800MHz bus into the masses. Intel’s plans for the next year don’t contain Pentium 4 models with a slower bus. The Quad Pumped Bus with 400 and 533MHz frequency will only remain within the value Celeron processor family.

Thus, the DDR400 dual-channel chipsets have been given the green light. Until the arrival of DDR II memory, scheduled for the middle of 2004, these chipsets will remain the top-end solution for the high-performance Pentium 4 platform.

Of course, DDR400 memory modules have caught the spotlight and now enjoy pretty high demand. Memory makers reacted to it by constantly increasing the supply. However, DDR400 modules currently in the market differ greatly in their characteristics. With this article we will try to share our recommendations on the approach to making the right choice about the PC memory. We will not dwell on particular modules for long, but will rather focus on general problems. The questions we are going to answer include:

  1. How do memory timings influence the overall performance of the Pentium 4 platform?
  2. Is it profitable to use high-speed overclocker memory when speeding up a Pentium 4 processor?
  3. What is more important for the Pentium 4 platform – higher memory frequency or lower latency?
  4. Are there any “ideal” memory modules?

Some time ago, we have carried out a similar investigation for the Athlon XP platform (see our article called DDR400 SDRAM with Athlon XP Platform: Performance and Future Potential). This time it is Pentium 4’s turn.

Testbed and Methods

Our testbed was built on an i865PE-based mainboard. We consider this chipset (and mainboards based on it) the most widespread and popular solution for the Pentium 4 platform with the 800MHz bus. Moreover, this mainstream chipset differs really slightly from the high-end i875P and we can extrapolate the results we will get on it to i875-based systems as well.

As for the particular mainboard, we chose ASUS P4P800 (see our Review of ASUS P4P800 Mainboard on i865PE Chipset for details). This mainboard provides good performance, and the latest versions of its BIOS are free from any stability issues, which may occur in some other mainboards with certain memory modules and/or memory subsystem settings.

Overall, the testbed looked as follows:

The testbed ran under control of Microsoft Windows XP Professional SP1. The BIOS Setup of the mainboard was set to maximum performance.

Influence of Memory Timings on Performance of Pentium 4 Platform

Our excursion around nearby computer shops brought the following news. Available DDR400 SDRAM modules have drastically different characteristics. You can find very fast modules supporting 2-2-2-5 timings (CAS Latency – RAS# to CAS# Delay – RAS# Precharge – Active to Precharge Delay) next to much slower ones. Many DDR400 modules currently available require the use of CAS Latency = 2.5, but this is not the worst variant. There eve exist some relatively cheap modules that need timings set to 3-4-4-8.

That is why we decided to benchmark our Pentium 4 3.2GHz (800MHz FSB) system with dual-channel DDR400 SDRAM using different timings. This test will give us some clues as to the influence of the memory characteristics on its speed. We will also see if it is really that advisable to search for high-quality DDR400 memory modules. In order to make the review more illustrative, we also include the results shown by a system with a slower processor – Pentium 4 3.0GHz. Moreover, we carried out a few tests using DDR333 SDRAM in the asynchronous mode (remember that this memory actually works at 320MHz in i865/i875 systems).

First of all, let us consider the results of synthetic benchmarks that measure the memory subsystem performance. The Cachemem test brought the following news:

 

Pentium 4 3.2
DDR320 | 2-2-2-5

Pentium 4 3.2
DDR400 | 3-4-4-8

Pentium 4 3.2
DDR400 | 2.5-4-4-8

Pentium 4 3.2
DDR400 | 2-4-4-8

Pentium 4 3.2
DDR400 | 2.5-3-3-6

Pentium 4 3.2
DDR400 | 2-3-3-6

Pentium 4 3.2
DDR400 | 2-2-2-5

Memory read speed, MB/s

2569.3

3452.5

3459.4

3484.1

3511.3

3515.9

3707.4

Memory write speed, MB/s

1141.4

1278.5

1279.2

1356.5

1364.9

1447.3

1601.6

Memory copy speed, MB/s

2093.3

2403.7

2405.5

2470.1

2540.4

2623.8

2939.4

Latency

362

289

288

283

281

276

242

There are several important points to discuss. First, according to Cachemem, the use of DDR320 (DDR333) memory with a processor supporting 800MHz FSB leads to a considerable slowdown of the memory subsystem. You can clearly see it from both: catastrophic bandwidth reduction as well as the significantly increased latency.

As for DDR400 that works synchronously with the system bus, the changes of its memory timings do not cause any significant change of the bandwidth. They mostly affect the latency. Note also that the CAS Latency parameter alone does not influence the memory subsystem speed too much. Two other parameters, namely RAS# to CAS# Delay and RAS# Precharge, affect the results much more heavily. This is a curious fact, considering that many memory sellers usually draw the customer’s attention to the CAS Latency parameter, often without even mentioning the other ones. So, we recommend you to note that modules with CL = 2.5 or 3 will not in all probability work with minimal RAS# to CAS# Delay and RAS# Precharge.

The next benchmark is SiSoft Sandra 2003:

It simply confirms the things we have said above: CAS Latency is not to be considered the main factor in determining the DDR400 memory speed in the Pentium 4 platform. In our tests, we have a clear example: the memory with timings set to 2.5-3-3-6 is faster than the memory with 2-4-4-8 timings.

The same is true for the results we got in the memory subsystem test from PCMark2002. As for the CPU Score, this value largely depends on the processor frequency, as this test doesn’t process any big data amounts that would require intensive memory access. However, I think we’ve had enough of synthetics, let us turn to real applications now.

The Business Winstone 2002 test measured the system performance in typical office applications like MS Word and Excel. However, the memory subsystem speed contributes a lot to the overall result in this benchmark. For example, we may note that Pentium 4 3.0GHz with faster memory turns to be a preferable solution over Pentium 4 3.2GHz with slower memory. This fact also suggests that you should see to the quality of the memory with no less attention than to the selection of the central processor.

Multimedia Content Creation Winstone 2003 tests the system performance in digital content creation tasks. This test includes applications for processing images, video and audio. As you see, we have dramatically different results here. Thus, Pentium 4 3.2GHz with the slowest memory outperforms Pentium 4 3.0GHz with DDR400 SDRAM and 2-2-2-5 timings. The fastest Pentium 4 3.0GHz platform only matches the performance of the Pentium 4 3.2GHz one when the latter uses DDR320 SDRAM in the asynchronous mode.

Data compression in the popular WinRAR program is a curious test. We see that the bandwidth and latency of the CPU-memory highway is crucial for the total result here. For example, Pentium 4 3.0GHz falls behind Pentium 4 3.2GHz by 3% only (when the same memory timings are used). Meanwhile, the difference between “good” and “bad” DDR400 SDRAM may reach 22%, in case all other system characteristics are equal.

Although video encoding into MPEG4 format also involves data compression, the results differ much from what we have seen in WinRAR. As Multimedia Content Creation Winstone 2003 showed us, that the CPU speed is an important factor for video encoding. Thus, Pentium 4 3.0GHz with the fastest memory always performs worse than the Pentium 4 3.2GHz platform with any memory, no matter how slow it is.

The final rendering of an image in professional applications is a task quite indifferent to the speed of the memory subsystem. It is the CPU speed that plays the decisive role here.

When shadows are rendered in professional applications, the memory performance is more important, although the CPU speed is still the critical factor.

The algorithms lighting calculations are, on the contrary, much more dependent on the memory performance. We have Pentium 4 3.0GHz with DDR400 (2-2-2-5 timings) losing only to the Pentium 4 3.2 platform with the same memory.

The popular 3DMark benchmark of the current and previous versions produces very similar results. Fast memory is always a boost for such gaming tests. It means that if you are assembling a gaming platform, you may want to choose your memory very carefully.

The CPU test from 3DMark03 shows it clearly, too. The Pentium 4 3.0GHz platform with DDR400 and with the minimal timings is always faster than the Pentium 4 3.2GHz one, but with less aggressive timings. Once again, we would like to note the fact that RAS# to CAS# Delay and RAS# Precharge parameters affect the system performance more than CAS Latency does.

Quake 3 definitely likes faster memory. As you can see in the diagram, we made it run 10% faster by simply reducing the memory timings.

The timings are less important for Unreal Tournament 2003. Nevertheless, aggressive memory settings help the Pentium 4 3.0GHz platform outperform the Pentium 4 3.2GHz one.

We see a very similar picture in Serious Sam 2: The Second Encounter gaming benchmark.

Overall, we can state that the selection of memory modules is a keystone of a platform featuring Pentium 4 processor and 800MHz system bus. When the quality (and speed) of DDR400 SDRAM does not match the processor, your Pentium 4 won’t show you everything it can. Across a number of tests, we witness a platform with slower memory losing in performance to a platform with slower processor but faster memory.

Overclocking Pentium 4 Processors and Overclocker Memory

Besides ordinary DDR400 SDRAM (often marked as PC3200), many manufacturers offer faster modules like DDR433 (PC3500), DDR466 (PC3700) or even DDR500 (PC4000). JEDEC, however, did not standardize any DDR SDRAM with a frequency above 400MHz. Anyway, many memory manufacturers learned to design memory chips supporting higher working frequencies than the standard 400MHz. It is exactly for the modules built of such chips that they use “DDR500” marking and the like.

The second part of this review is dedicated to these particular memory modules capable of working at more than 400MHz and their advantages, if there are any. The currently available dual-channel chipsets from the i875 and i865 families for the Pentium 4 processor do not provide any multipliers to increase the memory frequency. It means that you may need your memory to go faster than the regular DDR400 SDRAM only when you are up to some exciting CPU overclocking by means of increasing the FSB speed above the nominal 200MHz. On the other hand, i875 and i865 do provide diminishing coefficients (divisors) of 5:4 and 3:2, so you do not actually have to use an overclocker memory with an overclocked FSB. All in all, you have three options:

FSB frequency

Memory frequency

FSB:MEM=3:2

FSB:MEM=5:4

FSB:MEM=1:1

200 MHz

133 MHz (DDR266)

160 MHz (DDR320)

200 MHz (DDR400)

210 MHz

140 MHz (DDR280)

168 MHz (DDR336)

210 MHz (DDR420)

220 MHz

147 MHz (DDR293)

176 MHz (DDR352)

220 MHz (DDR440)

230 MHz

153 MHz (DDR306)

184 MHz (DDR368)

230 MHz (DDR460)

240 MHz

160 MHz (DDR320)

192 MHz (DDR384)

240 MHz (DDR480)

250 MHz

167 MHz (DDR333)

200 MHz (DDR400)

250 MHz (DDR500)

260 MHz

173 MHz (DDR346)

208 MHz (DDR416)

260 MHz (DDR520)

270 MHz

180 MHz (DDR360)

216 MHz (DDR432)

270 MHz (DDR540)

280 MHz

187 MHz (DDR373)

224 MHz (DDR448)

280 MHz (DDR560)

290 MHz

193 MHz (DDR386)

232 MHz (DDR464)

290 MHz (DDR580)

300 MHz

200 MHz (DDR400)

240 MHz (DDR480)

300 MHz (DDR600)

The first conclusion that comes to mind when we look at these results implies that using memory in synchronous mode with the overclocked FSB may bring some performance advantages. However, this may not be quite true. As we have already seen, memory timings play an important part in determining the overall performance of an i875/i865-based platform. And the overclocker memory  working at higher frequencies cannot boast aggressive timings. Moreover, the production technology of memory chips is such that chips intended for work with aggressive timings at 400MHz just cannot achieve (even if we use poor timings) the frequencies of overclocker chips, although the latter have weaker timings at 400MHz. For example, we have not yet encountered DDR500 modules that would work with 2-2-2-5 settings at 400MHz. On the other hand, however high-quality it may be, a DDR400 module cannot work stably at 500MHz, although it uses the most aggressive timings at 400MHz.

Thus, the overclocker is confronted with a problem, which of the two options to choose: overclocker memory with relatively poor timings, but working synchronously with the FSB, or a lower-frequency memory with aggressive timings, which requires a diminishing coefficient.

To carry out our test, we used a Pentium 4 2.4C processor that had been very good during overclocking (see our Intel Pentium 4 2.4C Overclocking article). We tested this CPU with the FSB sped increased up to 250MHz. In this mode, this CPU with a fixed multiplier (12x) works at 3GHz, while the overclocked FSB allows using memory at 250MHz (500MHz DDR), 200MHz (400MHz DDR) and 167MHz (333MHz).

We also had this processor working with 275MHz FSB, that is, at 3.3GHz clock-rate. Currently available memory modules do not support such FSB frequency in the synchronous mode, so we had only two options: 220MHz (440MHz DDR) and 168MHz (336MHz DDR). For each test, we picked up memory modules that would work with minimal possible timings for the given frequency. Thus, we tested DDR500 with 2.5-4-4-5 timings (DDR500 modules from GeIL, OCZ or Corsair may do), while DDR400 was used with 2-3-3-6 timings. In all other cases, we chose the most aggressive timing scheme: 2-2-2-5.

For a better comparison, we also showed the results of “true” (non-overclocked) Pentium 4 3.2 and 3.0GHz processors with the best DDR400 memory.

Following our good tradition, synthetic tests come first. The table below lists the results we got when measuring the memory subsystem performance in Cachemem.

 

Pentium 4 3.0 (FSB 1000MHz)
DDR333 | 2-2-2-5

Pentium 4 3.0 (FSB 800MHz)
DDR400 | 2-2-2-5

Pentium 4 3.3 (FSB 1100MHz)
DDR336 | 2-2-2-5

Pentium 4 3.0 (FSB 1000MHz)
DDR400 | 2-2-2-5

Pentium 4 3.3 (FSB 1100MHz)
DDR440 | 2-3-3-6

Pentium 4 3.0 (FSB 1000MHz)
DDR500 | 2.5-4-4-5

Memory read speed, MB/s

3232.8

3560.7

3567.7

3110.4

3694.8

4128.5

Memory write speed, MB/s

1278.2

1597.1

1330.2

1347.1

1580.9

1331.6

Memory copy speed, MB/s

2260.1

2931

2603.4

2773.8

2729.3

2793

Latency

290

226

283

250

267

200

You see that DDR500 SDRAM working synchronously with the 250MHz FSB is not unrivaled even in the synthetic tests. Although this combination works quite well, far not the best memory timings spoil the results in the data write and copy tests. The combination of DDR400 SDRAM with 2-2-2-5 timings and 200MHz FSB is good enough too and sometimes proves even better than DDR500. As for other competing solutions, they cannot perform as fast as DDR400 SDRAM with aggressive timings. We are going to see a little later how this affects the system performance in real applications. For now, let us run a few more synthetic benchmarks.

The performance in Sandra benchmark first of all depends on the bandwidth of the CPU-memory bus. That is why no wonder that Pentium 4 2.4C overclocked to 3GHz and supported by DDR500 memory did much better that the others. DDR400 takes the second place.

The memory subsystem test from PCMark2002 produces a slightly different picture: DDR400 SDRAM working in synchronous mode with the FSB yields better results than DDR440 working at 5:4 of the FSB frequency.

Now, let us turn to the promised real applications. Be ready for some surprises here.

Business Winstone 2002 seems to confirm the general rule. Processors with higher clock-rate provide higher performance. In case of the same CPU working frequencies, the platforms get ranked according to their memory speed. The only curious, but anyway quite predictable, result is that Pentium 4 2.4C overclocked to 3GHz and using DDR400 memory at 5:4 of the FSB frequency runs faster than the “true” Pentium 4 3.0 with DDR400 SDRAM in the synchronous mode. This is probably because the FSB frequency not only affects the overall system performance but also influences the processor bus bandwidth a lot.

Multimedia Content Creation Winstone 2003 has it somewhat another way. For example, Pentium 4 overclocked to 3.3GHz and working at 275MHz FSB, but with DDR336 memory, falls behind the “true” Pentium 3.2GHz with fast DDR400. It means that Content Creation Winstone 2003 needs the memory subsystem to work faster. This observation receives yet another proof from the fact that Pentium 4 overclocked to 3.0GHz, but using the diminishing coefficient for the memory frequency performs worse than the regular Pentium 4 3.0, notwithstanding the higher speed of the front-side bus.

WinRAR seems to be most sensitive to the memory subsystem performance among all the benchmarks we use. Sometimes it produces simply wondrous results, as we see here. For example, Pentium 4 3.0GHz and Pentium 4 3.2GHz working with DDR400 SDRAM with 2-2-2-5 timings outperformed all of their overclocked competitors, notwithstanding a considerable clock-rate gap. The reason is simple: when overclocking, we had to either worsen the timings or use the system and memory busses asynchronously. As you see, WinRAR is fastidious about that.

The numbers for the MPEG4 encoding task look more true-to-life. At least, the CPU frequency is the decisive factor here to determine the overall system performance. Note also that the platforms with Pentium 4 2.4C overclocked to 3GHz and DDR500 memory are running faster than their analogs that use DDR400 with the diminishing coefficient, notwithstanding the more aggressive timings in the latter case.

It is the CPU speed that determines the system performance during final rendering (ray tracing and shadowing) in professional packages like CINEMA 4.0. That is why we see nothing extraordinary here.

The nature of lighting algorithms is different and the memory speed is the most important thing for them. Thus, Pentium 4 2.4C overclocked to 3GHz and using DDR500 in the synchronous mode manages to outperform itself, overclocked to 3.3GHz and using asynchronously working memory.

The “true” Pentium 4 3.0 and 3.2GHz processors take the top places in 3DMark thanks to the high-speed DDR400 SDRAM with 2-2-2-5 timings present in their platforms.

Fps measurements in Quake 3 do not bring any revolution. There is only one fact to point out: when the FSB frequency is boosted to 275MHz, it is more advantageous to use 3:2 coefficient for the memory clock-rate rather than 5:4 in this test as well as in some others.

Unreal Tournament and Serious Sam 2 produce similar results. Any advantage in processor frequency leads to a higher fps value. The FSB speed and memory timings also contribute to the total result.

Summing it up, we should acknowledge that the memory tests we performed with the FSB overclocked to 250 and 257MHz did not bring any extraordinary insights. The only interesting thing from the practical point of view is that at 275MHz FSB it is often better to use 336MHz memory with aggressive timings rather than 440MHz memory with RAS# to CAS# Delay and RAS# Precharge set to 3. However, the performance gain is very small in this case anyway.

As for using DDR500 SDRAM in the synchronous mode during FSB overclocking to 250MHz, we really doubt that it makes sense. This overclocker memory is much more expensive than ordinary DDR400 SDRAM, while the performance it provides is just a little higher than that of DDR400 SDRAM with 2-2-2-5 timings. The table below shows what difference between these two combinations exists in various applications.

Benchmark

Performance gain with DDR500 SDRAM

Business Winstone 2002

0.6%

Multimedia Content Creation Winstone 2003

1.83%

Data Compression, WinRAR 3.0, Best, sec

2.83%

MPEG-4 Encoding, FlasK 0.78.39/DiVX 5.02, fps

0.67%

3DMark03, Default

0.33%

3DMark03, Default, CPU score

1.56%

3DMark2001 SE, Default

2.26%

Quake3 (four), High Quality, 1024x768x32

3.98%

Unreal Tournament 2003 (dm-antalus), 1024x768x32

3.95%

Serious Sam 2 (The Grand Cathedral), 1024x768x32

2.45%

CINEMA 4D, CINEBENCH 2003, Raytracing, CB

0%

CINEMA 4D, CINEBENCH 2003, Shading, CB

1.25%

CINEMA 4D, CINEBENCH 2003, Lighting, CB

1.75%

On average

1.8%

Memory Modules: Closer Look

It’s now time to show our hand and tell you about the memory modules we used in our tests. We will have an opportunity to discuss their operational characteristics later in this article, while now we just present a list of them. Two manufacturers, Corsair and OCZ, both the leading makers of fast and high-quality memory, had their products included into this review. So, we received the following modules for our tests:

A pair of 256MB modules that are supposed to work at 400MHz with 2-3-2-6 timings. In practice, these were the only modules to work at 400MHz with the lowest timings (2-2-2-5).

A pair of 256MB modules that are supposed to work at an up to 467MHz frequency. Corsair guarantees these modules to work at such a high frequency with 3-4-4-8 timings. Moreover, these modules are tested by the manufacturer specifically for stable functioning in dual-channel Pentium 4 platforms.

A pair of 256MB modules, positioned as budget memory for overclockers. According to the manufacturer, they support frequencies up to 467MHz with 3-4-4-8 timings and with DIMM voltage increased to 2.75V.

Two DDR467 modules from OCZ’s elite series. According to the manufacturer, they should work right at 466MHz with rather aggressive 2-3-3-7 timings and with 2.75V voltage. Regrettably, the practice showed that these modules work somewhat slower than they should.

Two DDR500 modules of 256MB capacity. They are claimed to work at 500MHz with 3-4-4-8 timings and 2.75V voltage. Our tests showed that they are actually operational with more aggressive timings. Moreover, their parameters turned to be better than even those of OCZ EL DDR PC-4000 Dual Channel Gold.

As you see, we have quite a motley assembly of modules, but it is the more interesting to test them at work.

So, we tried to use them with different memory timings. Selecting a timings scheme, we were increasing the frequency of the module to the stability peak. The DIMM voltage was set to 2.75V.

And we got quite curious results:

These are the most liberal timings and PC4000 from OCZ notches the highest frequency. Strange as it might seem, it is accompanied by Corsair’s PC3700, although the latter is not supposed to work with such timings at all.

After lowering CAS Latency by half a clock, we see that Corsair TWINX512-3700 goes down, while the other modules are more indifferent to this change.

With reduced RAS# to CAS# Delay and RAS# Precharge, the maximum frequency of Corsair TWINX512-3700 drops even lower. This time we have one more sufferer: OCZ EL DDR PC-4000 does not make it to 500MHz frequency.

When the CAS Latency dropped to 2, Corsair’s PC3700 refused to work at frequencies of 400MHz and above. The other modules kept on working fine.

Corsair TWINX512-3200LL modules were the only ones to handle such aggressive timings. Note however that this same memory was no record-breaker with non-aggressive timings.

This fact is yet another confirmation of the above-said rule: the memory intended for super-high frequencies does not like aggressive timings, even if the frequency it works at is much below the norm. On the contrary, the memory intended for work with minimal timings cannot support high frequencies. This way we have to admit that there is no universal memory for the PC. That is why when choosing your own memory, you should decide beforehand if you are going to overclock the memory bus in your system. If you are set on that, high-frequency overclocker memory may be the right choicer for you. Otherwise, you would be better off using memory with lower specified frequencies but supporting more aggressive timings.

Conclusion

In fact, we have already pronounced our inferences in the respective sections of the review. Now, let us put it all together:

We guess you have caught our point – selection of proper memory modules for your system is a critical and complex choice. We hope that our recommendations will help you in finding the right memory for your own outstandingly fast Pentium 4-based platform.