no performance gain when moving from CAS 3 to CAS2

[tiran]

look at http://www.techwarelabs.com/reviews/memory/memory_timings/index.shtml

(thanks to prof133 for the link.)

All i concluded from reading on the internet is that i will gain Very small gain when i invest on those "special" / "high speed" memory .
Am i right ?
I don't intend to overclock . got Intel 875 motherboard with P2.4 FSB800.
or in another words - have you measured a perfomance gain when you switched from CAS3 regular / slow memory to CAS2 HyperX etc memory ?
thank
tiran.

 

[bum_jcrules]

<A HREF="http://www.techwarelabs.com/reviews/memory/memory_timings/index.shtml" target="_new">Making it clickable.</A>


The article is good, but not great.

There are a lot of other things that they could have done with it. One question is why did they use an older chipset? Why did they go back to a 100MHz base signal and not start with 133MHz and then include a 200MHz signal? Those are questions that I will not get an answer to.


As for your motherboard tiran, you have a totally different scenario. First off you are using a totally different chipset. The i875P is known to not be friendly to more aggressive memory latency settings.

So all of the information from the Techwarelabs article is about 90% useless to to. Basically all you can do with it is to learn about different timing latencies.


As I stated before, the i875's MCH is not happy with tighter timings.

<A HREF="http://www.intel.com/technology/memory/ddr/valid/dimm_results.htm" target="_new">Here is a listing of the validated unbuffered memory modules for the i875.</A>

You will notice that for the CAS Latency settings the modules were between 2.5T (cycles="T") and 3T.

The best timing from any module was a 2.5T CL, 3T tRCD, and 3T tRP. The worst was 3-4-4.

One thing that you will not find on the Intel website is the configuration of the other latency settings. For most Intel chipsets it needs a CMD timing of 2T or higher. This can be set lower but you can start encountering stability problems. There are also Page Hit Limit (PH) problems with the i875 in Dual mode. In single mode the i875 can have 32 pages open but in dual mode it can only have 16 pages open at one time. This can kill performance in terms of missed pages. That in combination with a CMD set too low is brutal.

As for your question about "All i concluded from reading on the internet is that i will gain Very small gain when i invest on those "special" / "high speed" memory. Am i right?"

That is not entirely true. Memory frequency is still the most important factor in terms of performance gains within specific memory architectures. (RDRAM, DDRSDRAM, QDRSRAM, etc.)

The total performance gains from memory timings are tiny compared to the ones gained from frequency increases.

And for your "have you measured a performance gain when you switched from CAS3 regular / slow memory to CAS2 HyperX etc memory ?"

At higher frequency speeds, the increases from lower latency settings are diminished. This is because each cycle duration becomes shorter and shorter. 133MHz is 133 million cycles per second. 166MHz is 166 million cycles per second. So you can see that 33 million more cycles occur in the same timeframe and thus each individual cycle is taking less time to take place. And if the individual cycle is shorter so is the penalty from each latency cycle.


If you are not intending to overclock, standard settings are going to be more than sufficient for you. Just set your BIOS settings to "SPD". The SPD is a chip on your memory modules that contain the specifications of that module. SPD stands for Serial Presence Detect. The MCH sends out a signal and gets this information back from the modules. The BIOS then runs the memory at those specifications.




<A HREF="http://www.millionmanlan.com/MMLDefault.asp" target="_new">Gaming the THG way... THG sponsors the MML2: June 25-29, 2003 in Louisville Kentucky... Be there!</A>

 

[Guest]

There we go:

- assuming f = frequency in hz, t = time of a cycle in secs
- when f approaches infinity, t = 1/f = zero, therefore latency or clock cycle become near zero
- this means that as DDR speed goes up there is no point in reducing wait states using large cost for low-cl-ram.

Nows whos for a 4Ghz Intel P4 in a vapochill sandwich with single DDR400?

 

[bum_jcrules]

I don't see where you are going for sure... can you explain where you are headed with that comment?

Either you are saying that latencies will diminish to something near zero...

and/or...

You are saying that as frequency speeds increase there is no longer a need to reduce the latency settings because the clock cycle penalties are also diminished.


I will make one statement and then expound on it. "Physics is still physics." When a material is altered by Electro-magnetic stresses it will still require some time "x" to return to its original energy state.

This is the whole reason why latency cycles are needed... <A HREF="http://www.m-w.com/cgi-bin/dictionary?book=Dictionary&va=hysteresis" target="_new">Hysteresis</A>.

This applies to SDRAM, SRAM, RDRAM, NVRAM, etc. All have some form of a need for latency penalties because of the physical properties that make up the chips and modules themselves.

Once we hit the timeframe of "x” and continue to increase the frequency speeds, the number of clock cycles to achieve an acceptable ground state of the material will increase.

So as clock speeds increase... latency cycles will decrease to a point and then start to rise again if all other things remain constant.




<A HREF="http://www.millionmanlan.com/MMLDefault.asp" target="_new">Gaming the THG way... THG sponsors the MML2: June 25-29, 2003 in Louisville Kentucky... Be there!</A>

 

[Guest]

>you are saying that latencies will diminish to something near zero...

Yep and then we have the two and a half cycle, like inbetween 2 and 3.

 

[tabytha77]

Well I haven't even seen PC3200+ rated at above CL2.5, so the real choice is 2 or 2.5. The question for me would be whether to get 1 GB of 2.5 PC3200 or 512 MB of 2.0 PC3200 for the same price. I would think you could get better practical performance with double the RAM than faster latency.

 

[bum_jcrules]

Buy the 1GB. You will do better by having more capacity than a small difference in latency settings.

<A HREF="http://www.millionmanlan.com/MMLDefault.asp" target="_new">Gaming the THG way... THG sponsors the MML2: June 25-29, 2003 in Louisville Kentucky... Be there!</A>

 

[Guest]

Good decision...

Anyway these Hardware Manufacturers must have decided that Ram speed is the new benchmark.

Personally I think a good understanding on the difference between the use of a whole wait state to allow for "Hysteresis" and the use of a specification such as wait state of CL 2.5 ~ CL3 is whats important to remember. Whether you are pounding on the mouse button during a Lan meeting or pounding on the Bosses brain cells away from home.

If you're wondering what I think is on agenda nowadays, its a bit like a serial communications port where the RAM has a wire to say its "ready to send" and whilst this wire is in the "off" state the CPU can continue with other processes.

All speculative speaking of course or my hobby would not be computers.