Memory confusion

Do you have a question? Post it now! No Registration Necessary.  Now with pictures!

Threaded View

Can anyone recommend any articles about memory?  I want to buy two
sticks of 512MB PC3200 for a EPoX EP-5EPAJ board, and the prices (and
timings) are different.  For example,


I know the first number is CAS, and the lower numbers the better.
Some vendors only give the first number.    What is the significance
of these numbers, how will they affect performance, which ones run
cooler, and which one is a "best buy?"   TIA

Re: Memory confusion

Phisherman wrote:
Quoted text here. Click to load it

I wish CoolTechZone was still up, they have the best explanation I've
ever seen.

The best memory timings possible are 2-2-2-5, at least according to the
people who set the standard. Each of the numbers represents how long the
system has to wait for the memory to be in a ready state before data is
fetched or delivered or modified. So the higher any of those numbers,
the longer you have to wait for the data.

The first number is CAS (Column Address Strobe) latency, and if you only
see one number advertised for a stick of memory this is likely to be the
one, it is the most important.

The 2nd number is RAS-to-CAS delay, RAS stands for Row Address Strobe.
After that is RAS Precharge delay, and after that is Act-To-Precharge

So in the example 2-2-2-5

The memory has a CAS latency of 2, RAS to CAS latency of 2, RAS
Precharge latency of 2, and an Act To Precharge latency of 5.   This
would be a top of the line stick of memory. (Corsair XMS makes some fine
2-2-2-5 with very nice samsung chips, and if you are willing for a
slight bit longer wait on Act To Precharge - Crucial Ballistix makes an
excellent value 2-2-2-6 memory).

The difference between a memory timing of 2-2-2-5 and a more common
timing of 3-4-4-8 in most systems is about 25-30% in overall memory
system performance. It makes about a 15-20% difference in actual system
performance in my experience, so it is a signifigant upgrade if you can
afford it.  The two biggest bottlenecks in most modern computers systems
are the memory subsystem and the hard drives. Investment in these can
actually yield better resuls than spending the same amount of money on a
faster processor.

For instance, an Althlon 64 3500 system, the next jump up is a couple
hundred dollars more for maybe a 7% real world performance gain. But by
spending $100-110 more on the memory that goes into it, you can gain
15-20% easily.   So good memory is not a place to skimp if you are
trying to build a truly fast machine. Supposing you already had raid
support on your motherboard and you used the other hundred dollars on
another harddrive so you have have RAID-0, this in turn would also yield
a considerable performance benefit - at least 10% overall and probably
closer to 40% on disk intensive applications.   To get a 30% system gain
with pure processor power would certainly cost more if you started with
an Athlon 3500.

One note about high end memory. You usually have to losen the timings a
bit if you overclock the FSB, so you lose some of the performance
advantage there, but as a general rule the lower the timings the memory
supports the better it will overclock (with looser timings).

In answer to your specific question: The 2.5-3-3-6 is the best performer
of the memory you listed.


Re: Memory confusion

On Mon, 28 Mar 2005 13:55:58 GMT, Phisherman

Quoted text here. Click to load it

"Best buy" would of course depend on the prices, which you
failed to mention and might vary depending on
who/when/where/how-much you buy.

Generally a name-brand budget grade 2.5,3,3,6 is the best
buy.  Contrary to the previous poster's estimates about
performance differences most put the difference at closer to
10% performanace between slow and fast memory except in very
limited, specific apps that are memory bottlenecked.  In
most common uses the differences are even less, closer to
3%, but then for those uses a modern system is overkill too.

You mentioned the board but not the CPU.  A Celeron with
lower FSB and performance doesn't necessarily warrant
higher-end memory as much, and further, if you bought PC3200
and then ran it at PC2700 speed, the originally specs
timings may go lower @ PC2700 speed.  That is, PC3200
2.5,3,3,6 memory might run at 2,3,3,5 (or some other timing,
this is just an example) when the motherboard is set to this
lower MHz memory bus speed.  

Overall one often considers what's appropriate compared tot
he rest of the system... If you paid premium for a high-end
CPU and video card, drives, etc, then premium memory makes
more sense.  If the rest of the system was budget/value
price optimized then it makes less sense to blow out that
value by spending a disproportionate amount on memory.

Re: Memory confusion

kony wrote:
Quoted text here. Click to load it

Actually, a 'budget' system with onboard graphics that shares system
memory would see a bigger increase in performance in the graphics area
than one with a high end AGP card. The difference can be very impressive
indeed with onboard graphics systems in FPS and smoothness.

At very high resolutions & bit rates the biggest bottleneck is that the
AGP/PCI-E graphics card has saturated it's fill rate and you wont see a
lot of difference in a high end AGP/PCI-E card with that. If you get 1-2
fps worth of improvement you are doing very well, it's maxed and that is
pretty much all there is to that. At lower resolutions and bit rates
you'll see more of a difference as there is some room for system calls
to the memory controller to have an impact.  With onboard graphics the
difference is much more dramatic as it's calls are entirely dependant
upon the system bus & memory controller and this is a bottleneck for it,
whereas your high end AGP card wont be making anywhere near as many
calls to system memory(as it has it's own very high speed memory)and
wont see the types of gains for that reason. Though we are talking
nanoseconds of access time, we are talking about a lot of accesses in a
short amount of time.

CAS   System Bus Speed  Cycle Time     Real World Latency
2.5      200 MHz      5.0 ns      12.5 ns
4.0     333 MHz     3.0 ns             12.0 ns
4.0     400 Mhz         2.5 ns           8.0 ns
2.0     400 Mhz         2.5 ns           5.0 ns
4.0     500 Mhz         2.0 ns           8.0 ns

(There are flaws in this example, mainly that the other timings would be
equal in this example and it only shows advantage for CAS but I'm
figuring out how to measure RAS impact is far harder than I want to get
into. X-2-2-5-1T is used here, each of the other settings has it's own
impact as well.)

Even with 40% less bus speed the CAS 2.5 memory is nearly keeping up
with 4.0 CAS memory with that huge bus speed advantage. I should point
out though, that this is only for access. You can pump more data through
a faster bus.  So it doesn't mean the throughput is as good as faster
bus (it isn't quite there because you do get more burst with faster
bus), but you can greatly accelerate the loading and unloading process
by reducing the latency.

   Due to DMA the processor isn't the only thing which can access the
memory controller directly -- almost all of your devices can - network
cards, hard drives, sound cards, graphics cards & chips. Impacts on
different systems vary, but the entire system benefits from reduced
latency. Anything where memory is a bottleneck is going to benefit.

There is another number usually attached to memory timings (but you
almost always have to visit the manufacturers site to get) and that is
command time, which is generally represented as 1T and 2T. 1T means the
entire process basically takes 1 cycle to sync up, and 2T means it takes
2 cycles to sync up (load, call the row, time the row has to remain in
memory, time to flush it, time to sync).  So if a cycle takes 2.5 ns to
complete a 1T stick can do it in 1 cycle(2.5 ns) and a 2T stick will
take 5ns or 2 cycles to do it. This can have as much impact as CAS on
memory performance. Most brand name memory is 1T.

And as Kony stated, if your systems calls for PC2700 and you put in
PC3200, it may be able to run at more aggressive timings at the lower
speed. Even if you can not get the system to be stable with a CAS of 2
(or whatever is lower than the listed CAS), I would still try lowering
the RAS as you can get some measurable benefits there. Just make sure
you run some stability test & benchmark it. Sometimes you get lucky and
get a good set of sticks that run error free without the added expense.
More often it's trial & error testing setting the RAS & CAS with lower
rated sticks for optimal performance and stability.


Site Timeline