Original Link: https://www.anandtech.com/show/267
There's no feeling like soaking up a bright spring day with the wind blowing in your hair and the smell of fresh flowers and the world around you flowing crisply through your body. However then you reach work, school, or wherever you happen to go on a daily basis and everything is removed from your mind and you are struck by the reality of the situation, you're stuck in a place you'd rather replace with another scene for the next 8 hours. It's days like that which make us all want to get away from the terrors of living a repetitive life, but at the same time it brings us closer to the things we cherish in life the most. Our favorite things. To the tune of the Sound of Music we all find ourselves singing about our favorite things at one time or another, although it may not always be in the same musical fashion as Julie Andrews once graced the stage with, there is no denying that each and every one of us out there takes the time to appreciate those which we cherish and are fond of. Not necessarily in a materialistic sense, but we all have our favorite days, cars, smells, tastes, electronics and feelings. And although the combination of emotions and technology in that analogy may be a bit awkward, to any computer hardware enthusiast, it should feel right at home.
If you've ever known a true hardware enthusiast, or if you happen to be one yourself (chances are, if you're reading AnandTech, you are) you'll know the feeling you get when your benchmarks come out a few percent ahead of the competition, or when you make a wise investment in a product that'll last you a little longer than whatever else is out there. It is because of the extremely competitive nature of most computer hardware enthusiasts and tweakers out there that we find the need to have Message Forums and Newsgroups where hundreds upon thousands of us can get together and chat about the latest hardware, and how to squeeze that extra 3 frames per second out of our gaming rigs or how to improve those compile times by another 30 seconds. From a distance it may look like a sad sickness, however just as any hobbyist would rush to defend their time-gobbler, any computer hardware enthusiast would tell you that the love of the hobby is the drive behind it all, which isn't always the most practical reasoning as computer hardware can grow to be very expensive.
Unfortunately, with everyone out there concentrating on building the fastest overall system, there is very little attention paid to the stability of the system. The stability of any system comes from a number of factors, while we may all talk about using the most stable motherboards with the most efficient cooling systems, very little attention is paid to one of the most important parts of any computer system, the memory. Usually enthusiasts will prefer to get whatever memory modules happen to be available at the most reasonable price, and they'll pick up as much as they can. Because having 192MB of RAM sounds much more powerful than having 64MB of RAM, and it sounds like you'll actually be using your computer for important tasks rather than having a LAN party with your friends or playing multiplayer Unreal with a couple of tough bots, which you'll probably end up doing in the end. The bottom line is that very little attention is given to the quality of SDRAM modules on the market today, and because of this you see an influx of lower quality parts that some poorly educated users may end up falling for when they're constructing their systems. And when you're a student on a budget, or don't have the luxury of "fixing" a poor investment with another, hopefully, more intelligent one you'll have to stick with your first choice for as long as you possibly can before making that next upgrade, so it's wise to make the first choice the best one possible.
How Times Have Changed
AnandTech's last SDRAM comparison came about in September of 1998 where the goal was still to weed out the poorly manufactured modules from the high quality solutions offered for sale on the net. However the times have changed in the 6 months that have elapsed, more manufacturers have started offering their own branded SDRAM modules and vendors are cutting corners to edge one another out of a few bucks in the market. In doing this, one would think that the consumer would be the one to benefit, unfortunately the consumer isn't able to compare every single module out on the market before making a buying decision. When a vendor says that they don't have the particular modules you're looking for in stock but they have something that's "just as good" you either take their word for it or shop elsewhere, which isn't always the safest thing to do if you've never dealt with the vendor.
At the same time, more expensive usually means better, right? Well why is it that in some cases, the more stable modules are those that are actually $10 cheaper than the rest offered by the vendor? There are many happenings behind the scenes that most consumers aren't aware of that could potentially cost you considerably in the end with a memory purchase, luckily with the changing times are new ways to avoid getting ripped off, and as usual, AnandTech is ready with the explanations.
Let's start off with by defining the search for the best SDRAM memory module out there today…
Performance is not the issue
For those of you who are looking for the fastest PC100 SDRAM out there, it is about time to halt your search. Performance is not the issue when considering SDRAM, especially PC100 SDRAM. Regardless of what some benchmark programs can produce, in real world situations, a single, well-made PC100 SDRAM DIMM will provide virtually identical performance to any other DIMM of its class.
At the same time do not get fooled into the nanosecond hype that fell over the PC industry during the initial introduction of the elusive "10ns SDRAM DIMMs." The nanosecond rating on a particular DIMM module does not directly relate to the performance you'll be receiving from your system overall, so don't be conned into buying a more expensive module just because it is rated at a lower speed.
Stability
So if performance isn't the issue, then what is? Well, the true purpose of shelling out all that money for high quality SDRAM is to ensure stable system operation. At normal memory bus speeds (66/100MHz depending on your setup) most SDRAM modules will work just about the same as far as stability is concerned; simply because they were designed to run at those frequencies to the best of their ability while keeping costs down. If all you plan to do is run your system at the 100MHz Front Side Bus speed then any of the modules compared here will give you the stability you need, it is when you start to push the limits of your system that we can truly separate the best from the rest.
If you recall, there are two parts to every SDRAM equation, the PCB and the SDRAM chips. Both parts of this equation contribute equally to the overall quality of the final module. As a refresher, the PCB, or Printed Circuit Board, is the green board the circuitry of the SDRAM module (or any RAM module for that matter) is literally printed on. It is this printed circuit board you handle when you are installing your DIMMs and it is this board that dictates how clean the electrical signals flowing through your SDRAM will remain. Clarity of signal (i.e. the absence of noise in the current) is what determines stability from the PCB end, a well manufactured PCB will keep all trace lengths within the recommended values set forth by Intel, have a minimum of 6 layers in the construction of the PCB and will make use of EMI Suppression techniques. While you can't really ask your vendor if their SDRAM implements any of those basic requirements for PC100 classification, it is good to have some background information on what separates a well made PCB from one of lesser quality without getting a degree in Electrical Engineering, although at times, one wouldn't hurt.
The second part of this RAM equation are the SDRAM chips themselves. These are the chips that dictate how stable the final SDRAM module will perform at various bus speeds. The most popular chips are the widely used Samsung chips, however you will see chips from NEC, Fujitsu, Hyundai and Toshiba on other popular modules. Here's where you need to be careful as things can start to get a bit tricky with the ratings on these individual chips.
Price
Another issue that never came into question in the September comparison was the price of the individual modules, simply because there weren't any "novelty" entries into the comparison. However as previously mentioned, times have changed, and this time around AnandTech received some SDRAM samples that were a bit out of the reach of the average consumer, but included in the comparison nonetheless.
The price to stability ratio is a consideration that we must all make unless you happen to be in the rare position where cost isn't a factor at all. At the same time, you need to shrug any teachings that may have told you that a higher price means a product that requires higher quality material to construct and is therefore superior. Case in point would be the controversial Samsung G8 vs GH chips that have been floating around the storefronts of many vendors. While the G8 chips are rated at a higher clock speed, their actual stability is lower than that of the GH chips at higher Front Side Bus (FSB) frequencies, however the G8 modules are priced noticeably higher than their superior GH counterparts by some vendors. Why on Earth would something of lesser quality be priced higher than a better overall solution?
The driving force behind this unique situation happens to be the classic price war scenario, where one vendor sets the price on a particular product, and in order to compete with that vendor, another vendor selling an identical part will set a similar price on that product, even going to such extremes as making the price difference a mere dollar. In the end, the consumer is the one who benefits, however the mentality of most is that the more expensive solution is the better solution, when in reality, the more expensive solution was the result of a price war and not in fact the better solution. So lesson number two is to disregard everything you may have learned about price differences and remember that more expensive doesn't necessarily mean better.
CAS Latency
System RAM is accessed in rows and columns by other components in your system, if you look at 64MB of RAM as a table split into a number of cells then it becomes easier to understand how RAM is accessed. Although the actual process isn't as simple as opening up Excel and searching through a few data cells, it is the fundamental idea behind accessing RAM. Both of these strobes are signals that your CPU or other device (like the processor on your video card) sends to your RAM. They tell a circuit in your RAM module that an address line is correct. In the case of RAS, that the row is correct, or in the case of CAS, that the column is correct. The speed at which you access a single row of RAM is defined as the Row Access Strobe Latency, or RAS Latency; and as you may be able to guess, the speed at which you access a single column of RAM is defined as the Column Access Strobe Latency, or CAS Latency.
CAS Latency ratings vary from chip to chip, and also varies depending on the bus speed used with the modules. While a chip may carry a CAS rating of 2 (lower is better) at 66MHz [bus speed] the rating on the chip may rise to 3 if used at 100MHz. Meaning that if you were to set the CAS Latency via your BIOS Setup to a value of 2, then upped the FSB frequency to 100MHz you would be going beyond the specification of the chips themselves and you could start to experience instability. The ideal, yet realistic, goal is to get a module that is rated at CAS 2 for 100MHz and CAS 3 for anything higher than that, be sure to get the facts straight from your vendor before placing an order as this will ultimately dictate how stable your RAM will operate at higher bus speeds.
Very Few True PC100 Modules
If you remember the hype at the introduction of Intel's PC100 Memory Specification, you'll remember that Intel wanted to increase the difficulty of poor quality modules to make their way into the hands of the consumers by implementing a strict identification policy for all PC100 memory modules. While the idea was sound in theory, it lacked the enforcement necessary to make this bill a law, and the industry went on pretty much as it had in the past. Fortunately the market did gain some stability out of the whole PC100 compliance fad of 1997 in that many manufacturers started increasing the standards for the production of their memory modules. Although the improvements were significant, they were still no where near as strict as what Intel had originally hoped to accomplish with the PC100 standard.
As a refresher, the PC100 specification was briefly outlined by the following 11 calls to action:
- 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
Interpreting the Specification
Ok, its great that we know what the PC100 Specification includes, however what does that matter if we have no idea why in the world specifications for trace width and spacing must be precise, and what placing series termination resistors on the data lines allows us to do? So let's filter out some of the garbage and come up with a "translated" version of the PC100 specification.
One of the biggest problems with the generic SDRAM modules you can go out to just about any vendor and pick up was that their modules, while they could be using high quality NEC chips, were built on low quality Printed Circuit Boards (PCB's). Think of it as putting a Ford engine in a Porsche frame, don't expect to get the same performance as if you used a genuine Porsche V-8, the same concept applies to SDRAM, although it is an obscure comparison it relates almost seamlessly. For the most part, generic SDRAM "sticks" as they are often called, used high quality chips from companies like NEC, Micron, Hitachi, or Samsung (SEC), however the PCB's were manufactured in a cost effective manner, in most cases rendering the SDRAM modules incapacitated when used in certain combinations. In order to cut costs some manufacturers chose to use 4-layer PCB's compared to the recommended 6 and 8-layer PCB's for SDRAM, as a result 4-layer boards were much susceptible to physical damage, and as you might guess a 4-layer board contains much less electrical insulation compared to 6 and 8-layer boards, this provides for increased levels of noise in the current. For this reason the PC100 specification clearly states that modules must be manufactured on 6-layer PCB's, and nothing less, so say good bye to those cheaper 4-layer boards.
Since the introduction of the PC100 specification, AnandTech has not run into a single SDRAM module that has experienced stability or incompatibility problems during normal operation as the result of a poorly manufactured PCB. The once common 4-layer boards are almost extinct with the remaining few being those belonging to the PC66 era, with the higher quality 6 and 8-layer boards dominating the marketplace. Once again, the user is the one that benefits.
Try this little experiment, take a normally sharpened pencil, and using a straight edge draw a vertical line. Moving the straight edge a millimeter or so to the left draw another line parallel to the first one, now take a look at what you've done. You have two lines, however the lines aren't completely independent of one another, provided your pencil wasn't razor sharp the lines probably started to blend into each other somewhat in between the two.
The same thing applies to SDRAM, the PC100 specification includes explicit details as to the width and spacing of trace lines (the lines printed on PCB's) in order to minimize the levels of cross talk (electrical interference, on the experiment above it would be the points where the two lines blended into each other) between traces adjacent or parallel to one another. The SDRAM modules out on the market today don't follow any set specification for trace length and width, some manufacturers chose to use specifications much like those described in the PC100 specification, while others are far from that.
There are two more parts to the PC100 specification that must be mentioned in order to get the main idea of the purpose of the specification, the first is the Detailed SDRAM component specification. This basically states that a manufacturer MUST use PC100 compliant SDRAMs (the actual chips) in order for their modules, which also must meet the PC100 specification, in order for them to market their modules as PC100 compliant SDRAM. The chips used on PC100 compliant SDRAM modules must be manufactured using a special die and must have an access time of around 8ns.
Finally, the PC100 specification provides a detailed description of the timing parameters and data to be included on the onboard Serial Presence Detect (SPD) EEPROM located on ALL PC100 SDRAM modules. This SPD EEPROM is used to communicate with the motherboard's chipset (e.g. 440BX, 440ZX), and although it won't really make a difference with Aladdin V, MVP3, and SiS 5591 motherboards, it allows the BIOS and the Chipset to communicate with the memory to properly configure the memory timings to control the memory. Without a properly programmed EEPROM the BX chipset may misinterpret an instruction or the BIOS may "guess" as to how the SDRAM should be configured, in both cases resulting in unstable system operation.
One small addition included in the PC100 specification is the requirement for a standard identification on all PC100 SDRAMs, the labels are required to have the format PC100-abc-def, where the 'abc-def' string would be replaced by 4 key timing parameters of the module and the revision number of the SPD specification. The label idea never really caught the momentum Intel would have wanted it to, as very few manufacturers actually use the PC100 labels on their modules at all. Out of all of the SDRAM AnandTech received for evaluation, two modules actually had the PC100 label on them, the rest were plain with the exception of their manufacturer's sticker prominently placed on them. |
The Myth: PC133
Although some manufacturers did submit "PC133" compliant SDRAM modules, the specification has yet to be released from Intel, and all modules claiming to be "PC133 compliant" are actually good guesses at what the specification will call for. Chances are that once the true PC133 modules hit the market, the bar of stability will be raised once again, however until then, don't fall for any manufacturers claiming to have PC133 modules already available. In fact, AnandTech's showed that the preliminary PC133 samples didn't even outperform some of the standard PC100 modules in terms of stability.
In time, with the push from Intel and AMD to push the limits of the standard FSB frequency to 133MHz and beyond, the PC133 specification should become a reality, however until Intel officially announces anything, there is no PC133 compliance officially in existence. Without an official specification, what may be the specification today may be different from what the specification states a week from now. All manufacturers are doing now is guessing at what the final specification will be and are hoping to get a head start on the competition by having PC133 compliant modules ready and waiting at the release of the specification.
As you've probably already learned, in the computer hardware industry, patience is one of your greatest virtues, don't jump the gun too quickly with PC133. Let the manufacturers sort themselves out, then make your move.
How Much RAM is Enough?
We've all asked this question at one point or another, but how much RAM is enough? The answer is entirely situation specific, however you can get a general idea on how much RAM you'll need using a quick benchmark comparison.
Today's business applications notice almost no improvement over 64MB of RAM, although there is a measurable improvement all the way up to 512MB of RAM, a difference of 0.2 business Winstone points is not noticeable at all to any sort of user. Business users will find that 64MB of RAM under Windows 98 is enough.
Using S3's Monster 2 Quake 2 demo, which contains an enormous amount of textures, running at 1600 x 1200, on an AGP TNT based accelerator, which in theory should require a heavy amount of system memory accesses to swap the large textures in and out of, the most noticeable jump is that experienced from 64MB to 128MB. Making a clear indication of a future movement to 128MB as a requirement, however you can get by with 64MB, even as a gamer.
The Naturally Speaking Voice Recognition benchmark shows an incredible improvement when going from 64MB of RAM to 128MB, however after the initial boost in performance, the drop off is significant, illustrating that the realistic needs for even a professional user running Windows 98 remain at the 128MB point.
Encoding using the Microsoft Netshow benchmark provided an almost indiscernible difference among the different memory configurations, even 64MB is enough in this case.
Adobe Photoshop is an extreme memory hog, especially when you're dealing with manipulations of 20MB+ images. The sweet spot for a Windows 98 system here is 192MB, the drop in render time from 64MB to 128MB is considerable, however the drop from 128MB to 192MB is much more noticeable. All in all, the sweet spot for a Windows 98 system seems to be between 64MB and 128MB, with the general trend of requirements taking the ideal Windows 98 PC in the direction of a 128MB minimum setup.
The picture changes ever so slightly under Windows NT 4, a more robust (memory hog) operating system in comparison to Windows 98. There is a noticeable difference between 64MB and 128MB of RAM under Windows NT, however anything above 128MB of RAM provides a series of diminishing returns, with 256MB being the absolute peak before the benefits lose themselves in the added cost of the memory upgrade.
In most multi threaded applications, such as those high end developing and image editing suites (e.g. Microsoft Visual Studio, Adobe Photoshop), the jump from 64MB to 128MB and then to 256MB is noticeable. Most NT workstations will want a minimum of 128MB of RAM, with 256MB being the absolute sweet spot. Although having 384MB of system memory can be nice, it isn't necessary, and the performance gain experienced in going from 128MB to 256MB is nearly twice that experienced in going from 256MB to 384MB. In the future things may change, however by today's standards, for most NT workstations, anywhere between 128MB and 256MB is ideal.
Preparing the Test
In order to isolate the memory modules as the only realistic causes of any fluctuation in stability, choosing the proper test bed was a bit of an ordeal, luckily AnandTech was aided in lab by the wonderful folks over at Kryotech who supplied AnandTech with a room temperature cooling system a week before the first stability tests were to commence. At the heart of AnandTech's SDRAM stability test bed was a Pentium II 333, capable of being reliably overclocked to 416MHz, running at room temperature with the aid of Kryotech's Renegade ATX-PE Room Temperature Cooler. The ambient case temperature of the Renegade test bed was kept at room temperature, or approximately 22 degrees Celsius, as was the surface temperature of the Pentium II processor.
The 333MHz Pentium II was chosen for its versatility in terms of clock multiplier support, as AnandTech's sample remained clock unlocked, and allowed for the usage of the 2.5x clock multiplier when testing higher FSB settings. In order to gain support for a wide variety of FSB frequencies, two motherboards were used as the basis for the test bed, the choice to use two motherboards came to make sure that there were no specific incompatibilities between the SDRAM being compared and a particular motherboard.
The entire test bed was configured as follows:
CPU | Intel Pentium II 333 |
Motherboard | ABIT BX6 Revision 2 & AOpen AX6BC |
Video | Matrox Millennium II PCI |
Hard Disk | Western Digital 5.1GB Caviar UltraATA |
Operating System | Microsoft Windows 98 |
The stability was evaluated according to the percentage of complete test runs (a complete test run is one that runs crash-free) the module managed to pass at each individual FSB frequency. The tests consisted of AnandTech's stability tests, which were continuous trial runs of Business Winstone 99, each trial run requiring that all three parts of the test complete successfully before passing.
The Candidates
The requirements for submission were as follows: All entries into the March 1999 AnandTech SDRAM Memory Comparison were to be made in pairs, and in configurations of 8 x 64 (64MB), 16 x 64 (128MB), and 32 x 64 (256MB). The modules must be available for sale within three months of the release of the article. So without further ado, we have AnandTech's top SDRAM candidates. (click images to enlarge)
Azzo PC133 Pre-Release Module | 64MB |
64MB |
Corsair SEC Modules | 256MB |
128MB |
|
64MB |
Enhanced Memory Systems PC133 HSDRAM | 128MB |
128MB |
|
64MB |
|
64MB |
Goldstar LGS | 64MB |
64MB |
Memman Mosel Vitalic | 128MB |
64MB |
Memman SEC w/ 3rd Party PCB | 64MB |
Memman SEC Originals w/ SEC PCB | 64MB |
64MB |
Mushkin Hyundai | 128MB |
128MB |
Mushkin SEC w/ 3rd Party PCB | 128MB |
128MB |
|
64MB |
|
64MB |
Mushkin SEC Originals w/ SEC PCB | 64MB |
64MB |
Mushkin SIEMENS | 64MB |
64MB |
The Results
100MHz FSB | 112MHz FSB | 124MHz FSB | ||||
CAS - 2 | CAS - 3 | CAS - 2 | CAS - 3 | CAS - 2 | CAS - 3 | |
Azzo PC133 | 100% | 100% | 100% | 100% | 100% | 100% |
Corsair SEC | 100% | 100% | 100% | 100% | 86% | 100% |
EDRAM PC133 HSDRAM | 100% | 100% | 100% | 100% | 100% | 100% |
Goldstar LGS | 100% | 100% | 100% | 100% | 100% | 100% |
Memman Mosel Vitalic | 100% | 100% | 100% | 100% | 0% | 0% |
Memman SEC1 | 100% | 100% | 100% | 100% | 86% | 100% |
Memman SEC Originals2 | 100% | 100% | 100% | 100% | 100% | 100% |
Mushkin Hyundai | 100% | 100% | 100% | 100% | 80% | 100% |
Mushkin SEC1 | 100% | 100% | 100% | 100% | 86% | 100% |
Mushkin SEC Originals2 | 100% | 100% | 100% | 100% | 100% | 100% |
Mushkin SIEMENS2 | 100% | 100% | 100% | 100% | 80% | 100% |
129MHz FSB | 133MHz FSB | |||
CAS - 2 | CAS - 3 | CAS - 2 | CAS - 3 | |
Azzo PC133 | 0% | 100% | 0% | 0% |
Corsair SEC | 0% | 100% | 0% | 0% |
EDRAM PC133 HSDRAM | 0% | 100% | 0% | 100% |
Goldstar LGS | 0% | 66% | 0% | 0% |
Memman Mosel Vitalic | 0% | 0% | 0% | 0% |
Memman SEC1 | 0% | 100% | 0% | 0% |
Memman SEC Originals2 | 0% | 100% | 0% | 86% |
Mushkin Hyundai | 0% | 86% | 0% | 0% |
Mushkin SEC1 | 0% | 100% | 0% | 0% |
Mushkin SEC Originals2 | 0% | 100% | 0% | 86% |
Mushkin SIEMENS2 | 0% | 86% | 0% | 0% |
1 - Third Party PCB
2 - Samsung (SEC) Made PCB
Although the most stable RAM in the comparison goes to Enhanced Memory Systems for their PC133 HSDRAM (High performance SDRAM), the lack of general availability of their HSDRAM modules (which are the closest things to above a PC133 specification) forfeits the victory to another tie between Mushkin and Memman with their Samsung (SEC) Originals modules. It seems like once again, the PCB makes all the difference, where the SEC chips with the SEC PCB manages to come out on top in all cases over the SEC modules with the third party PCB.
Who should you buy from, Memman or Mushkin? They are both reliable vendors, however unless you're buying more than just SDRAM, go for whoever happens to be cheaper...guys, let the price wars begin. Don't settle for anything less than a Samsung Original module, remember, the entire module must be manufactured by Samsung in order for it to be a SEC original (PCB & chips).
Close runners up are Corsair with their SEC based modules, however the word is that Corsair will be dropping their SEC based line in favor of Micron based modules. Their motives are still in the dark, however AnandTech hopes to find out shortly how the Micron parts turn out in the comparison. It seems like an investment in the Memman or Mushkin SEC Originals now, will last you the longest time until the PC133 specification is finally complete.
The modules to stay away from are the Memman Mosel Vitalic if you plan on running at above 112MHz, the stress endurance tests run by AnandTech proved the Mosel Vitalic to be unworthy of any sort of overclocked state. What you must keep in mind for all of these modules is that at 100MHz, there is no telling them apart, and even at 112MHz there is no real way of discriminating among the various SDRAM modules, so if you don't plan on overclocking, then as long as you get a decent PC100 module you should be fine.
Corsair was the only company to supply AnandTech with registered 256MB DIMMs, and as AnandTech's tests showed, the use of 256MB DIMMs was quite erratic in most motherboards. Most systems experienced quite a few cases of registry errors/corruption when the 256MB modules were installed, so if you're planning on installing more than 256MB of RAM, either check with your motherboard manufacturer first or purchase smaller sized DIMMs.
But when it comes down to AnandTech's favorites, the Memman and Mushkin Samsung Originals are the number one pick. Find the lowest price either of them will offer you, and go for it. Don't you just love price wars? Here's where you, the consumer, finally benefit.