Need advise on memory timings (Barton on Nforce II Ultra400 dual channel)

[Nil Einne]

Got a new system and currently working out what to run my memory at.
I'm mainly doing stability testing at the moment but having trouble
working out what to aim for and what to avoid. I thought it was
mostly, the lower the better but now know I was mistaken so need so
advise.

I have a DFI Infinity Ultra board which has the following settings:
T(RAS) = Active to Precharge delay
T(RCD-Read) = RAS to CAS delay for reads
T(RCD-Write) = RAS to CAS delay for writes
T(RP) RAS Precharge delay
CAS latency
T(RC) = bank cycle time
T(RPC) = refresh row cycle time

I have discovered from Mushkin that T(RAS) should never be lower then
T(RCD)+CAS+2 cycles for stability and performance. This was actually
quite significant since I seemed to be able to get reliable MemTest86
operation at outrageously low settings (3), but no one else had done
this and my memory has problems doing CAS latency of 2.0 at 200mhz so
it can't be super good so I realised there must be something wrong.

However, one thing I would like to know is is there any harm in
running RCD-Read and RCD-Write async other then potential stability
issues. Is there any advantage in running RCD-Write lower since from
what I can gather RCD-Read is what is important (I am able to run
RCD-Write at 2 but RCD-Read needs to be 3)?

Also, is there any reason I would want to adjust RC and RPC?

And could someone give some info on what will really help and in what
situations? I've come across numerous reviews put they all seem to
conflict each other. Some suggest CAS latency is the most important
but other people have found CAS 2.0 is unnessary on Nforce2 since
CAS2.5 is nearly the same.

Finally, is SiSoftSandra useful in determing performance (so I can
analyse e.g. a trade off of reducing FSB but improving timings)? Does
it really cover the wide range of demanding apps I'm likely to use
like video encoding, games etc? I don't believe it's a matter of pure
memory bandwidth since from what I've read there are other factors
that come into play as well.

 

[Ius Tull]

Finally, is SiSoftSandra useful in determing performance (so I can
analyse e.g. a trade off of reducing FSB but improving timings)? Does
it really cover the wide range of demanding apps I'm likely to use
like video encoding, games etc? I don't believe it's a matter of pure
memory bandwidth since from what I've read there are other factors
that come into play as well.

To check memory performance under different timings use

Sciencemark 2.0 - Memorytest / Freeware 270 kb

http://www.sciencemark.org/

Sometimes they are temporarily down, then use this link for download.

http://www.quickbench.mynetcologne.de/Be_memorysoft.htm

screenshots here

http://www.quickbench.mynetcologne.de/B_sciencemark.htm

hth

 

[Skid]

One of the best single sources for an explanation of these settings is the
bios guide at www.rojakpot.com

Many Nforce users will tell you that an even higher TRAS is helpful, in the
9-11 range. Perhaps someone could elaborate here.

The settings in any particular system are largely a matter of trial and
error, there are too many variables in brands and types of memory, even
stick-to-stick variations in matched pairs. Throw in voltages and temps,
varying software loads, and basically you're on your own.

The best app for determining memory stability is MemTest86, freeware that
runs from a boot floppy. Just punch in the timings you want to try in the
bios, then boot to MemTest and run a few loops. If it errors out you can
loosen the timings one at a time until you figure out the fastest stable
combo.

In Windows, Prime95 performs much the same function, with the added benefit
of stressing the cpu and testing your cooling.

Sandra will tell you which settings give you peak bandwidth; the other two
will tell you whether the fastest settings are also reliable. You should
also test Sandra in unbuffered mode (there are 9 settings to uncheck before
running the test, see the memory holy grail articles at www.anandtech.com
for directions.) The results can be surprisingly different and are generally
a better indicator of how much difference the settings actually make in
everyday performance.

 

[Nil Einne]

One of the best single sources for an explanation of these settings is the
bios guide at www.rojakpot.com

Many Nforce users will tell you that an even higher TRAS is helpful, in the
9-11 range. Perhaps someone could elaborate here.

The settings in any particular system are largely a matter of trial and
error, there are too many variables in brands and types of memory, even
stick-to-stick variations in matched pairs. Throw in voltages and temps,
varying software loads, and basically you're on your own.

The best app for determining memory stability is MemTest86, freeware that
runs from a boot floppy. Just punch in the timings you want to try in the
bios, then boot to MemTest and run a few loops. If it errors out you can
loosen the timings one at a time until you figure out the fastest stable
combo.

I know how to test stability. But for example, as I said, T(RAS) of 3
is stable but is worse performing. So what's important to me is are
there any other settings like this? I believe CAS latency etc are all
better or the same if lower so with these settings I might as well go
as low as I can. Obviously, if it comes to trade offs, e.g. I can go
at CAS 2.5 at 210 or CAS 2.0 at 200 I'll probably end up benching
which is why I asked for recommendations and I thank you for your help
there. Of course, any general ideas will be useful

P.S. I'll check out rojakpot soon

 

[Nil Einne]

One of the best single sources for an explanation of these settings is the
bios guide at www.rojakpot.com

Okay had a look now. It's useful but doesn't have
T(RC) = bank cycle time
T(RPC) = refresh row cycle time
yet (at least not in the free version)

I suspect they're not that important but would be nice to know

Sandra will tell you which settings give you peak bandwidth; the other two
will tell you whether the fastest settings are also reliable. You should
also test Sandra in unbuffered mode (there are 9 settings to uncheck before
running the test, see the memory holy grail articles at www.anandtech.com
for directions.) The results can be surprisingly different and are generally
a better indicator of how much difference the settings actually make in
everyday performance.

I knew it had some benchmarks of that sort. But what I'm not clear on
(and the Anandtech article didn't say either) was is peak bandwidth a
certain measure of performance as in if peak bandwidth is higher, apps
will either perform better or the same regardless of latency etc or
can latency etc affect apps depending on other factors. For example,
will random memory access be affected by latency settings more then
peak bandwidth (not that I'm totally sure if this will really be a
factor in my case)? The range of apps which really use the processor
I'm likely to run is quite wide and includes video encoding and
decoding (with features to improve quality), compression, checksum
calculation, games, distributed computing apps etc. I thought it might
be useful if these apps are affected to make a few diff settings since
my bios has CMOS reloaded so that I can change settings if necessary
when I'm planning to do something with a certain app for a long time.
Of course, it probably won't get used that much but maybe on occasion.

Having said that, my initial analysis tells me there is only one
setting which from what I've read, is usually inconsquential to
performance, that will force me to reduce FSB for better timings. My
RAM seem to have a limit in terms of timings and FSB which is the same
(At least for FSB 200 and higher). Vmodding doesn't help much either.
Still this might be useful to someone in the future...