If money were no object in buying a new computer system, enthusiasts would almost always choose the fastest memory with the best timings and the largest overclocking potential. The reality for most, however, is that a new system purchase or upgrade is most often a question of where to put your money for the greatest performance return. To make those kinds of decisions on a system purchase, you really need to know the real difference in performance between buying the best and buying something half the price, and possibly putting the money saved elsewhere for a larger performance gain.

These questions of the real impact of memory speed and timings on performance are things that we have talked about in many memory reviews. AnandTech has always been an advocate of real world performance measurements, and we've shunned using just synthetic benchmarks in our testing of every type of component. This is not because synthetic benchmarks are not useful - they are often very revealing of component differences. It is because running just synthetic benchmarks can severely distort the picture of performance with real applications and real games. That is why we always use games and the pure number-crunching Super Pi in our memory tests. It is also the reason why we test using both Buffered (Standard) and Unbuffered synthetic benchmarks. We have found in much of our testing that the less commonly used Unbuffered benchmarks more closely mirror how games really respond to memory differences.

This quest for real performance differences is also the reason why we moved to testing different memory speeds at the same CPU clock speed in our Athlon 64 memory tests. The AMD CPU, with unlocked multipliers, allowed us to finally remove the CPU speed differences from our memory tests so that you could finally see the true impact of memory speed increases and memory timings on performance. As you have seen in past reviews, those performance differences are very real, although they are much smaller than many memory manufacturers might want you to believe. On the other hand, faster memory speeds and faster memory timings do improve performance, no matter what some nay-sayers are determined to prove.

Let's state, right up front, that if you're on a tight budget, memory is often a good place to save a little money by buying less than the best and moving that cash to a better CPU or a better video card. You are likely to get a bigger performance boost with the extra $150 that you save on memory by buying a faster CPU or an upgraded video card. But don't be misled. The upgrade to the next higher CPU would normally increase performance maybe 5 to 10% - about the same as the difference in performance between cheap, slow memory and fast memory with tight timings. Video cards usually yield more than a 10% boost for that same $150.

If you do need to balance and buy less than the best in memory, the need for help is even greater than if you can afford whatever you want. Some Value RAM is pretty basic and slow, and some is just about as good as the best and most expensive that you can buy. For that reason, it was well beyond time for our first Value RAM roundup. There are some disappointments and a few very pleasant surprises. We think that you will be surprised by how much performance there is in some cases, and how little in others, when you're shopping for Value RAM.

The Memories
POST A COMMENT

102 Comments

View All Comments

  • Wesley Fink - Tuesday, April 12, 2005 - link

    xsilver -

    PLEASE read my Final Words. I made a very clear distinction between Value RAM for normal motherboards and Value RAM for Mad Overcockers. If I take your approach then it assumes there are no overclockers who buy a cheap motherboard, or a DFI nF4 Ultra for $134, and use cheap ram they push to the limit. I assure you there are lots of overclockers who push cheap ram to the limits, and who are looking for some great overclocking RAM at a cheap price. YOUR definition of Value is just one of many and we talk about this in depth on page 2.

    Most really cheap motherboards in the Value category have no adjustments at all for memory voltage, so I should logically only test at 2.5V or 2.6V using your approach. We are testing what this Value RAM can do with a consistent test bed.

    As for the 6800 Ultra and DFI, they were ONLY used so we could compare results to many of the high-end RAM we have recently tested - to keep the test bed consistent. I can imagine the uproar if I had tested with a 6200 and then listed past results with our 6800 Ultra test bed. We are not suggesting a Value System USE a 6800 Ultra, it is merely to have a consistent test bed.
    Reply
  • Wesley Fink - Tuesday, April 12, 2005 - link

    #40 - You are correct about overhead with Super Pi, but since we use a consistent test platform to test memory the Super Pi results can be accurately compared with each other.

    Your approach to running Super Pi in a RAMdisk is interesting, and should remove some barriers to broader comparison of CPUs and platforms.
    Reply
  • xsilver - Tuesday, April 12, 2005 - link

    cygni
    I've been reading anandtech since 1999
    and from that time there have been very few instances where CLEAR errors have been made.... I am afraid though this is one of them. I was trying to make a point that it is dissapointing as it is the first review right after anand did his whole spiel about integrity of reviews

    about corsair not co-operating... how about source the memory from somewhere else? newegg or other stores would be happy to oblige (they are a major sponsor... and you only need to borrow the ram)
    its just sad when corsair smells that you're doing a roundup and they know they could possibly look bad, so therefore they decide not to submit anything for review
    I mean imagine if there was a whole nvidia and ati shootout and one team know beforehand that they would lose... and they refuse to submit a card.... would people hear about if the card performs bad? or how companies are trying to do a shifty? people would be all over the forums spewing froth from their mouths!!!

    I consider anandtech the best tech website out there and in efforts to making it better I offer constructive criticisms... I don't know where the "All these "anand must be embarrased" posts every time are getting FREAKING OLD""
    is coming from but its the first that i've heard of it......

    and I don't know how to make it any clearer but this was a VALUE roundup right? who in their right mind would spend craploads of money on a dfi nforce4 + 6800ultra and get VALUE ram? .... I'm not saying you can't use good hardware.... just dont push it to 3.4 damn volts!!!!!

    Reply
  • Wesley Fink - Tuesday, April 12, 2005 - link

    #36 - On the AMD Athlon 64, there is no FSB since everything is derived from the HT bus. You are therfore correct that lowering the ram multiplier and running at DDR400 at a higher CPU speed makes little difference on the A64 platform. There are, however, two concerns with this approach:

    1) There is an Asynchronous Latency penalty, which can be tweaked somewhat on boards with better BIOS options like the DFI. It is not, hoever, the kind of asynchronous penalty you see on a FSB board like Intel.

    2) RAM multipliers are usually limited. If you have a standard set of 400, 320, and 266 speeds, you could only achieve DDR400 speed at a CPU frequency of 250. Anything lower than that would be running the RAM at less than 400. Most A64 CPUs can't do 250 on air at stock multipliers (the low end ones can) so they will be running less than optimum ram speed. That's where you could lower CPU multipliers or use a board like the DFI with lots of intermediate RAM ratios.

    In the end, overclocking on the AMD is a balancing act with more variables than the simpler FSB platforms. You are balancing LDT (the HyperTransport Multiplier which controls HT speed), CPU frequency, and RAM frequency to find the best mix for performance. You also have all the different memory controllers since they are on the CPU.

    In my experience there are not any simple rules for this, except faster values usually perform better than slower on the same variable. In other words, the easiest way to consistently improve memory performance is 1:1 memory speed.
    Reply
  • AnnihilatorX - Tuesday, April 12, 2005 - link

    Replying #29 Wesley Fink

    That's exactly a Gold series but NOT value series.
    Based on part number OCZ4001024WV3DC-K
    http://www.atacom.com/program/print_html_new.cgi?&...
    is the right one.
    See the reply below

    Replying #21 dvinnen

    I asked some other people on other forums. I've post some info on anandtech forum
    http://forums.anandtech.com/messageview.aspx?catid...
    Reply
  • highlandsun - Tuesday, April 12, 2005 - link

    Just a tiny point, I see this misperception all over - Super Pi is not a pure number crunching program. It writes huge sets of temporary files to disk and reads them back between loop iterations. A lot of people quote Super Pi times and assume they're all reasonably comparable measures of CPU+memory performance, but they're overlooking the fact that there's a significant chunk of I/O in there too, so disk type and filesystem state (fragmentation, etc.) will also impact the timing results.

    If you can install a ~128MB RAMdisk and run from there, then you can be sure that you're only measuring CPU+memory with Super Pi. (Then of course, the performance will depend on how good the RAMdisk's block memory move implementation is. But if you use the same version for all tests, you'll get consistent, comparable results.)
    Reply
  • ChineseDemocracyGNR - Tuesday, April 12, 2005 - link

    It's been mentioned "if you don't overclock..."

    You don't need expensive RAM to overclock the Athlon 64. On this platform the memory is always running on a divider, there's no performance penalty from overclocking the processor by increasing the reference clock but keeping the memory happy at DDR400.

    As for the $195 "Value" RAM, I wouldn't recommend it. Save $100 and get a more powerful video card or whatever else your system needs. This money could mean an upgrade from a GeForce 6200 to a 6600GT, or from a 6600GT to a X800XL; there's no RAM that makes up for that.
    You can achieve the same overclock on your processor with DDR400 RAM that doesn't like to be overclocked!

    I know this article was a lot of work, but I think it would be a good idea to have at least one other test platform. The DFI LP nF4 is known to be... "special" when it comes to memory compatibility. The results could be (and probably would be) very different on another board, like the more value-oriented nForce4 or K8T890 boards that would fit in better with this article. Heck, I think even the DFI NF4-DAGF would be a different story.
    Reply
  • Cygni - Tuesday, April 12, 2005 - link

    xsilver

    1) ALL reviews work this way. You think Autoweek BUYS the cars they review? You think PC Gamer BUYS the games they review? Anandtech asked for samples, and was given them. Or they are given to AnandTech to promote the product. That is how hardware reviews / review sites WORK.

    2) They were pushing the RAM to the absolute limits, which meant using voltages not available on most baords. Big deal! They also noted performance at each different speed setting, so even if your board doesnt have that voltage, they STILL had you covered.

    All these "anand must be embarrased" posts every time are getting FREAKING OLD. If you HONESTLY think that, there is NO WAY you have been here for more than a few months. Let alone years.

    Googer

    RTFA, and RTFC. For the love of christ...
    Reply
  • Googer - Tuesday, April 12, 2005 - link

    How can you do an article like this and NOT benchmark the KING of ALL Value RAM- Crucial? Reply
  • cryptonomicon - Tuesday, April 12, 2005 - link

    hmm I asked last review and I received I suppose...>> FINALLY << a comparison against BH5.. and by coincidence I'ts amazing that they are producing some, however it is just from those old dies lying around.

    as a side note, wth happened to kingston and mushkin. their stuff only hit 204mhz??? what a joke.

    :spelling corrections:
    "OCZ PC3200 Gold is sold as a 1 GB kit with two 412MB DIMMs at a price of about $195. "
    -change to 512mb
    Reply

Log in

Don't have an account? Sign up now