An entire year has passed since we last rounded up the latest SDRAM memory modules available online for a head to head comparison, and in the past year quite a bit has changed.

Our last memory roundup took place during the reign of Intel’s BX chipset as the platform of choice for most users.  The introduction of the BX chipset brought the need for SDRAM that could actually run at the 100MHz FSB frequency the chipset provided.  Prior to the introduction of the BX chipset there was little order established in the SDRAM market because of the incredible variation between the designs from one module to the next. 

Since motherboard manufacturers couldn’t possibly test their boards with very single memory module available on the market, they had to rely on following the motherboard design guides and the limited number of SDRAM modules they could get their hands on before their board’s release in order to make sure all the compatibility issues were worked out. 

Intel attempted to remove these SDRAM compatibility issues by creating their own PC100 SDRAM specification which outlined, in a very detailed manner, the requirements for a memory module to be classified as PC100 which would make it fit for use on a BX motherboard running at the 100MHz FSB. 

Although the PC100 specification wasn’t followed down to the last dot, most manufacturers followed it closely enough that stories of SDRAM compatibility issues were relatively scarce among BX motherboard owners. 

At the time of our last PC100 SDRAM Roundup we weren’t sure as to what the future of SDRAM would be because Intel’s “Camino” chipset hadn’t hit the streets yet.  One thing was very clear, by the end of 1999 there would be a definite shift towards 133MHz FSB platforms, but what type of memory would they use?

Since it turned out that the mysterious “Camino,” now known as the i820 chipset, wasn’t a good match for SDRAM at all and with Intel’s heavy push away from SDRAM and towards RDRAM it became quite clear that Intel wasn’t going to be the first to establish a PC133 memory specification, or even establish one at all.

Not willing to pursue RDRAM as a memory option for their chipsets, not at this stage at least, VIA stuck to SDRAM as the memory of choice for their 133MHz FSB VIA Apollo Pro 133/133A platforms.  But in order for VIA to take full advantage of the 133MHz FSB frequency their chipset would require an equally fast pathway to the memory which gave birth to the need for the 133MHz memory bus. 

Since Intel refused to acknowledge the need for PC133 SDRAM, after all their 133MHz FSB chipset solutions relied on RDRAM or PC100 SDRAM, VIA was put in a position where they had to take the initiative to create a standard if they wanted PC133 SDRAM to be adopted by the industry.  So, preceding the release of VIA’s Apollo Pro 133A chipset was the announcement that VIA had been working with the leading DRAM manufacturers on developing a PC133 memory standard.

And PC133 SDRAM is here to stay, for this year at least.  While the rest of the market transitions towards higher bandwidth solutions such as DDR SDRAM and RDRAM, the fact that PC133 SDRAM is pin compatible with current motherboards gives it a major advantage over more expensive emerging technologies.  At this year’s Spring IDF, Intel stated that they were finally going to be revising the PC100 specification for PC133 SDRAM in preparation for their first PC133 chipset solution, the i815 which is due out towards the end of Q2-2000. 

The platform we actually used for our tests was the VIA Apollo Pro 133A based ASUS P3V4X.  One of the beauties of the 133A chipset is that it can run your memory bus at the FSB frequency minus the PCI bus frequency, which in the case of a 133MHz FSB – 33MHz PCI bus would result in a 100MHz operating frequency for your memory bus which is what the PC100 specification calls for.

With a feature like that, why would you possibly need memory that can run at 133MHz or beyond?  Is the performance increase that significant?

There’s only one way to answer that question, let’s take a look at some benchmarks comparing a 100MHz memory bus to a 133MHz memory bus in the same exact system:

Even in Content Creation applications that are normally influenced very little by memory performance because of their dependency on a fast L2 cache are experiencing the PC133 craze as the CC Winstone 2000 scores indicate a 4% increase in performance just by switching to PC133 SDRAM.

Quake III experiences a 5% improvement in performance which is about the same performance increase you’d get from stepping up to the next processor speed in any current processor family (i.e. Pentium III 800 to 866). 

The SPECviewperf tests illustrate up to a 6% improvement in performance caused by the 133MHz memory bus.  While this may not seem like much, considering that the added price premium of PC133 SDRAM over PC100 SDRAM right now is negligible if you’re building a new computer there’s no reason not to go with PC133 if your motherboard can take advantage of it.

Then again, the performance improvement isn’t enough to go out and replace your current 128MB+ of PC100 SDRAM with PC133 SDRAM.

What makes a good Module

Now that we’ve established that there is a need for PC133 SDRAM let’s take a quick look at what makes a good module in the first place.

When Intel introduced the PC100 SDRAM specification in 1998 they put together a list of issues that needed to be addressed by both the chip and module manufacturers in order for the specification to be successful.  That list included the following:

• Minimum and maximum trace lengths for all signals on the module
• Precise specifications for trace width and spacing
• 6 layer PCB's with unbroken power and ground planes
• Detailed specifications for the distances between each circuit board layer
• Precisely matched clock trace lengths, as well as routing, loading, and termination requirements
• Series termination resistors on the data lines
• Detailed SDRAM component specification
• Detailed EEPROM programming specification
• Special Marking Requirements
• ElectroMagnetic Interference Suppression
• Selectively gold plated printed circuit boards

The actual specification goes into great depth for each one of those points, spelling out  exactly what the manufacturer must do in order to properly meet the specification.  Ideally this works out just fine because as long as you have a manufacturer that follows these specifications perfectly all SDRAM should be created equal and there shouldn’t be any question as to exactly how far your SDRAM will go when overclocking.

Unfortunately, the fact of the matter is that most of the SDRAM you can go out and buy features chips made by one manufacturer, a PCB made by another manufacturer, and is assembled at a third party plant. 

This brings out the point we have made in our past SDRAM roundups, that there are two parts to the equation used to make a quality DIMM: the SDRAM chips themselves and the PCB.  Both parts of this equation carry an equal weight, meaning that if you have some very high performing and ultra reliable SDRAM chips on a poorly manufactured PCB your module is most likely not living up to its true potential.

This is the reason that you can find SDRAM modules with identical labels on the individual chips themselves, but that can reach completely different frequencies simply because of differences in the manufacturing of their PCBs and the trace layouts on the boards as well. 

Since heat does not play a major factor in the overclockability of SDRAM modules, it is generally safe to say that although your mileage may vary from one module to the next, it isn’t as likely to vary as something like the overclocking potential of a CPU. 

CAS 2 vs CAS 3

If you remember back to our PC100 SDRAM Comparison from March 1999 we had a couple modules that could actually hit the 133MHz mark with reasonable success, so why is there even a need for newer PC133 SDRAM?

Other than the obvious fact that most PC100 SDRAM couldn’t even reach the 129MHz mark much less the 133MHz level, the modules that managed to reach 133MHz did so at a CAS Latency of 3 clocks.  The CAS Latency, or Column Access Strobe Latency helps to define how quickly memory can be accessed, and having a lower CAS Latency value should theoretically increase performance.  But by how much?

In order to answer that question we ran a few benchmarks illustrating the performance difference between running PC133 SDRAM at 133MHz/CAS 2 and 133MHz/CAS 3 on our ASUS P3V4X VIA Apollo Pro 133A test platform.

We get less than a 2% increase in performance in CC Winstone 2000 when decreasing the memory timings to CAS 2 on our PC133 setup.  This is mainly because Content Creation applications tend to reside within the L2 cache of a processor, and in the case of our Pentium III test CPU, the full speed on-die 256KB of L2 cache is enough for most of the CC Winstone applications.

Quake III Arena exhibits a 3% increase in performance when moving to CAS 2, it’s not a tremendous increase but it is something.

The SPECviewperf results show us that simply moving from CAS 3 down to a quicker CAS 2 memory timing can result in an increase of 1 – 13% depending on the situation, making CAS 2 PC133 SDRAM something that professional users should be very interested in as a 13% performance improvement is fairly significant. 

For most business/SOHO users having CAS 2 PC133 SDRAM isn’t of extreme importance in terms of performance, although gamers may appreciate the few extra frames it gives them. 

The Candidates

For this roundup we took a look at a total of 8 new PC133 SDRAM chips and tossed in an old contender in order to give you all a point of reference.  The candidates are:  Corsair PC133 using Infineon chips, Corsair PC133 using Toshiba chips, Memman PC133 using Infineon chips, Memman PC133 using Mosel Vitalic chips, Mushkin Budget PC133 using Nanya chips, Mushkin High Performance PC133 using Mosel Vitalic chips (2 revisions), Mushkin High Performance PC133 using NEC chips, and for a point of reference we tossed in a Samsung Original module using their -GA chips which was Samsung’s first SDRAM chip rated at 133MHz.

All of the modules tested were 128MB modules, except for Mushkin's NEC module which was a 256MB part.

Corsair PC133 - Infineon Memman PC133 - Infineon	Click to Enlarge
Corsair PC133 - Toshiba	Click to Enlarge
Memman PC133 - Mosel Vitalic	Click to Enlarge
Mushkin Budget PC133 - Nanya	Click to Enlarge
Mushkin PC133 - Mosel Vitalic Revision 1.0	Click to Enlarge
Mushkin PC133 - Mosel Vitalic Revision 2.0	Click to Enlarge
Mushkin PC133 - NEC - 256MB	Click to Enlarge
Samsung Original - GA	Click to Enlarge

The only note that needs to be made is that during the testing period Mushkin supplied us with a new revision of their High Performance PC133 SDRAM using the Mosel Vitalic chips, this revision we called revision 2.0 in order to separate it from the similar looking earlier revision. The only difference between the two seems to be that the newer revision has more fine tuned traces since it did turn out to be a better performing module.

The Test

There are a number of ways one can test for memory stability, including using low level memory testing software as well as hardware.  Considering that the majority of AnandTech’s readers are interested in the real world stability of these modules we decided to take a real world testing approach to this roundup.

We constructed an application test suite split into four categories: Content Creation (Multitasking), Content Creation (Data Intensive), Gaming, and High End 3D.  We selected one benchmark for each category and ran a loop of 10 complete tests and recorded the number of failures during that testing process.  If a setup completed 95% or more of the test runs without crashing then it was issued a Passed rating, otherwise it was given a Failed rating.  The reasoning behind this is that it isn’t ok for your system to crash more than 5% of the time under extreme situations.  It isn’t ok for your system to crash 5% of the time either but we are taking into account factors such as software related crashes, and other non-memory related errors.

For the Content Creation (Multitasking) tests we used Content Creation Winstone 2000, for Content Creation (Data Intensive) we used SYSMark 2000, for Gaming we used Quake III Arena, and for High End 3D we used SPECviewperf with each one of the benchmarks looped through an entire test run ten consecutive times.

We used an unlocked engineeing sample Pentium III so that we could keep the final clock speed of the setup, even when using an overclocked FSB frequency, below the rated frequency of the CPU thus eliminating the CPU as a potential cause of any crashes. All tests were run under Windows NT 4.0 because of its superior stability in comparison to Windows 98SE.

We chose the ASUS P3V4X as our 133A motherboard platform because of its numerous overclocked FSB settings which we could use to better test the modules. We tested at the following FSB/memory frequencies: 133MHz, 140MHz, 144MHz, 146MHz and 148MHz. No module made it above 148MHz on our test bed.

133MHz - CAS 3	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	YES	Pass	Pass	Pass	Pass
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	YES	Pass	Pass	Pass	Pass

 

140MHz - CAS 3	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	YES	Pass	Pass	Pass	Pass
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	YES	Pass	Pass	Pass	Pass

 

144MHz - CAS 3	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	YES	Pass	Pass	Pass	Pass
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	YES	Pass	Pass	Pass	Pass

146MHz - CAS 3	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	YES	Pass	Pass	Pass	Pass
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	YES	Pass	Pass	Pass	Pass

Even the old Samsung -GA module is able to keep up with the rest of the pack at 146MHz/CAS 3.

148MHz - CAS 3	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	NO	-	-	-	-
Corsair PC133 (Toshiba)	NO	-	-	-	-
Memman (Infineon)	NO	-	-	-	-
Memman (Mosel Vitalic)	YES	Fail	Fail	Fail	Fail
Mushkin Budget (Nanya)	NO	-	-	-	-
Mushkin High Performance (NEC - 256MB)	YES	Fail	Fail	Fail	Fail
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Fail	Fail	Fail	Fail
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Fail	Fail	Fail	Fail
Samsung Original (GA)	NO	-	-	-	-

Although all of the Mosel Vitalic based PC133 modules and the Mushkin 256MB NEC module POSTed at 148MHz, not a single one was able to complete any of our tests at that speed. It seems like 146MHz is the ceiling for all of these modules.

133MHz - CAS 2	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	NO	-	-	-	-
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	NO	-	-	-	-
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	YES	Fail	Fail	Fail	Fail

We are finally starting to see some failure. Although our old Samsung Original -GA is able to boot, it doesn't make it much farther.

And we finally understand what makes Mushkin's Budget memory cheaper than the rest, it won't work at 133MHz/CAS 2.

140MHz - CAS 2	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	NO	-	-	-	-
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	NO	-	-	-	-
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	NO	-	-	-	-

 

144MHz - CAS 2	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	NO	-	-	-	-
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	NO	-	-	-	-
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	NO	-	-	-	-

146MHz - CAS 2	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	YES	Pass	Pass	Pass	Pass
Corsair PC133 (Toshiba)	NO	-	-	-	-
Memman (Infineon)	YES	Pass	Pass	Pass	Pass
Memman (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin Budget (Nanya)	NO	-	-	-	-
Mushkin High Performance (NEC - 256MB)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 1 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Pass	Pass	Pass	Pass
Samsung Original (GA)	NO	-	-	-	-

Nothing really changes from 140MHz to 146MHz.

148MHz - CAS 2	POST	Content Creation Winstone 2000	SYSMark 2000	Quake III Arena	SPECviewperf
Corsair (Infineon)	NO	-	-	-	-
Corsair PC133 (Toshiba)	NO	-	-	-	-
Memman (Infineon)	NO	-	-	-	-
Memman (Mosel Vitalic)	NO	-	-	-	-
Mushkin Budget (Nanya)	NO	-	-	-	-
Mushkin High Performance (NEC - 256MB)	YES	Fail	Fail	Fail	Fail
Mushkin High Performance Revision 1 (Mosel Vitalic)	NO	-	-	-	-
Mushkin High Performance Revision 2 (Mosel Vitalic)	YES	Fail	Fail	Fail	Fail
Samsung Original (GA)	NO	-	-	-	-

And once again, nothing is capable of making it through our tests at 148MHz although the latest revision of the Mosel Vitalic based memory from Mushkin is at least able to POST as is their 256MB NEC module.

Conclusion

The first thing that must be understood is that although we attempted to place similarly performing modules on an equal level, not all modules are in fact created that way. There are errors that due occur that don't result in system halting crashes or loss of data, and to most AnandTech readers, especially the average business/SOHO user or gamer, these memory errors go undetected. If you were running a more mission critical setup, for example a webserver, these errors would most likely end up in the worst case scenario, server downtime, but considering that you'd never want to overclock anything that would be used as a full time server this shouldn't really apply.

In terms of the best bang for your buck, Mushkin's Budget PC133 SDRAM is fairly affordable at $99 and as long as you don't plan on using a CAS 2 setting (as it didn't work at CAS 2 at all) it should be fine up to around 146MHz.

For those users that are looking for a reliable CAS 2 capable module, the Infineon based Corsair/Memman modules worked out fine up to 146MHz as well as the Mosel Vitalic modules from Mushkin and the 256MB NEC module also from Mushkin.

For the most part, a well designed PC133 module featuring 7.5ns chips as did all of the modules in this roundup, should be able to handle the 133MHz setting without any problems and considering that most memory manufacturers are lenient with the -ns ratings on their chips, you shouldn't have too much of a problem pushing them another 5% - 10% above their intended 133MHz operating frequency.

As far as our pick out of the 9 in this roundup goes, the results speak for themselves, Mushkin's latest revision of their High Performance PC133 SDRAM with Mosel Vitalic chips comes out on top. As with any company, be careful when ordering however, because unless you make it very clear that you want this particular module when ordering, you may not get what you expected.

Links

Corsair - http://www.corsairmicro.com

Memman - http://www.memman.com

Mushkin - http://www.mushkin.com

Samsung - http://www.samsung.com