|
Comment #1
Hello,
I read with great interest your article on L2 Cache Latency testing. I have two questions. First, do you think that the results might be somewhat different if the tested processor were say a PII or PIII instead of a Celeron? Is there not some significant differences in the construction and operation of L2 Cache between the PIIs and PIIIs as compared to the Celerons?
Second, I recently had use of a SOYO 6BA+III, which supposedly had L2 Latency Adjustability, however any time I changed this setting from the designated selection, CPUID indicated the L2 had been disabled and overall machine performance suffered as if it had. Do you know if there is a problem with SOYO's
implementation of L2 Latency Adjustability or was something else going
on? During my period of use of the SOYO board I had both a PII-333 running at 450 and a PIII-450 running at 558 and saw the same results with both processors. Memory although fleeting, seems to bring up something I had read once about SOYO having a problem in this area but I failed to be able to find any reference.
Thanks for your time and efforts,
Ronald Keller
Hello Ronald,
I think there will be differences between the L2 cache latency
performances of the Celeron, the Pentium II and Pentium III. Because the
L2 caches of the Pentium II and III are external and run slower than the
Celeron, changes in the performance of the L2 cache is more likely to
bring about a bigger effect on overall system performance.
If the L2 cache has really been disabled by your adjustment, you will
be able to feel the drop in performance. A simple processor benchmark
can verify if the L2 cache was really disabled or not. If the processor
did not suffer any drastic drop in performance, you can be assured that
the L2 cache has not been disabled.
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #2
Adrian:
May I suggest doing the L2 cache latency tests again.... this time with a
different overclocked processor (how about the popular 366@550?) and different
RAM (the popular 128MB PC133?) to see if the results are specific to
that processor and RAM combination.
Alan R. Mege
Hello Alan,
Unfortunately, I do not have the resources to buy another processor
and more RAM just for such tests. However, I believe the results are not
processor-speed or RAM-grade specific. No matter how fast the processor
run, the effect of different cache latencies should remain the same for
the same processor as nothing fundamental has changed, merely its speed.
As for PC133 SDRAM, it will have virtually no effect at all if used
with the Celeron 366 @ 550MHz. That's because the SDRAM will run
synchronously with the system bus which is 100MHz for the Celeron 366 @
550MHz. So, it will be like running standard PC100 SDRAM. Also, because
we are testing only the performance differences between different L2
cache latencies, RAM performance won't affect that. Whatever performance
boost from running the SDRAM at a higher speed will affect all the L2
cache latencies. So, while the test results may differ, the difference
between the performance of the different L2 cache latencies will remain
the same.
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #3
Hello,
What does the "default" L2 cache latency setting in the BIOS set the latency
to, for Celerons (300A)?
Any help is much appreciated. Keep up the good job on your site :)
Thanks,
Arpad Sooky
As I mentioned in the Dissection
section, the default L2 cache latency for the Celeron is 5.
Hope that helps you some! :)
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #4
You may want to try the tests on a P3 or P2, since the cache is actually
separated from the processor core and that is where the real problem lies in overclockers. Celerons rarely have the same
hassles as P3 or P2's. If you don't have the processor to do it, let me know, I have a P3 450 and a BE6,
PC133 SDRAM, Quantum HD, Viper 770 32MB non-Ultra, SBlive! to do the same tests over for
you.... just send me exact procedures you used to assure validity.
Rick Dumont
Former writer for The Tech-squad (but EverQuest killed my time)
Wow! Thanks for the offer of help, Rick! That would be great since I
doubt I'll ever save up enough money to buy a Pentium II processor, much
less a Pentium III.
Overall, the test procedures involve keeping all system variables
equal except for the L2 cache latency. So, that means no changes
to the system must happen during the entire benchmarking process. You
will have to fully optimize your hard disk and then defragment all empty
space. Then after each benchmark run, you will have to defragment the
empty space on your hard disk again to ensure that the next test will
have the same environment.
All in all, you will need to run all benchmarks three times for each
and every L2 cache latency. Needless to say, that's going to take a
long, long time! ;-) The benchmarks I ran
were WinBench 99 1.1's CPUMark 99 and FPU WinMark 99
tests, 3D WinBench 99 1.1's 3D Transformation and Lighting and 3D
Transformation tests and all 3DMark 99 Max tests. If possible, try
to get benchmark results from CacheChk 6.0 as well. I ran all
those benchmarks but I still have yet to evaluate the data from CacheChk
6.0.
I would recommend running one round of all the benchmarks before
going to the second round. That's better than running the same benchmark
three times before going to the next one because this method will ensure
that the conditions will be equal for the tests. Hmm... I guess that's
about it. Hope you will have some results for us soon! :)
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #5
Regarding your article on L2 cache latencies.. I think you've left out
two important topics. Addressing these would make your benchmarks much more significant. Now that we know changing the L2 cache doesn't have a
significant performance impact...
(1) Does altering the L2 cache latency really improve your ability to get
a stable overclock? And if so, how much of a difference does it make?
(2) Since you tested with a Celeron (which has a different cache arrangement then the P2/P3) do the same results apply similarly to the
P2/P3?
Overall I was glad that someone wrote concerning this topic --- it had
been something I was curious about. I have a P3-450 that overclocks really well, but only with my L2 cache latency tweaked up. I was curious
how this was affecting my performance, and now I know. THAT IS, assuming
your findings carry over to the P2/P3 with its large half-speed cache :)
Thanks,
Darren Shepard
Hello Darren,
It's hard to say if changing the L2 cache latency will improve your
ability to get a stable overclock. Theoretically, it's certainly
possible but just like you saw in this article, theory isn't everything
it's made out to be. There's no way to find out without more tests which
would need to involve multiple "stubborn" processors which I,
most unfortunately, do not have. :(
Again, I believe the results will be different for the Pentium II and
III. I suspect the performance differences with those two processors
will be much more significant that it is with the Celeron.
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #6
Hello,
I have read your article on L2 cache latency. The results are very strange to my opinion. I don't doubt them however! Maybe this is the
solution for this problem: at lower latency settings more errors occur. These error however are corrected by the ECC-codes. This way we are not
aware of them. But we get a speed-penalty if we have a ECC-correction. I don't know of this all is true.... but it is an interesting
thought.
Roel Vliegen
PS English is not my first language... 8)
Hello Roel,
That's certainly a very interesting thought there! I didn't even
think of it. Thanks for giving me that knock on the head. I can
certainly visualise an update in the future which will include test
results for the L2 cache latencies with ECC correction disabled. Then we
will know if ECC correction really affects the performance of the
various L2 cache latencies.
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #7
Hi,
Just a question regarding the L-2 latency timing. Do you reach the same results when running your CPU at default clockspeed?
Regards,
Philip Leenen
Hello Philip,
I have not tested the L2 cache latencies at the default clockspeed
because the benchmarking of 15 different L2 cache latencies using 5
different tests (each with 3 runs) requires an extraordinarily long
amount of time. However, I'm confident that the performance differences
between the various L2 cache latencies will be the same, irrespective of
the processor clockspeed.
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #8
Good article, great idea. Just a few comments, when you're talking about
0.2% differences in a comprehensive benchmark like 3D Mark and Winbench, you
really should be doing at least 3 runs and average the results. Heck, I run
Winbench multiple times with the same system same latency and I get different results every time.
Other than that, I think you should explain what exactly it means to have a
latency of "14" or "2". Also, I get the feeling that the latency will be
different for everyone. What you would probably want is latency that corresponds to the CPU clock cycles, so that they are in sync. I suspect
that "13" is the worst and just one more, "14" is the best, is because 14 is
exactly sync'd whereas 13 is off by one and thus the CPU has to actually
wait "26" before it gets to read from the cache. Anyways.. good article,
hope you can make more improvements in the future too!
Bryant Fong
Hello Bryant,
As per standard benchmarking procedure, I ran all test three times
and the results were averaged before being presented in this article. I
also checked all benchmark results to ensure that they were consistent.
So, I'm quite sure the results are as accurate as they can get.
To be frank, I, too, suspected that it was a synchronization issue as
I mentioned in my Conclusion. But
as another reader has pointed out, there may
yet be another answer for the benchmark results - ECC Error
Correction.
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #9
Hi Adrian,
Hey, thanks for running those L2 latency tests, I'm sure it was something
everyone was wondering about. However, I think it would've been useful if a bit of the scientific method were
introduced into the tests regarding the benchmark programs used.
If you simply ran 3DMark 15 times at the SAME L2 setting, I have a suspicion
that you would find a variance of around 1.7%. This would imply that the
benchmarks aren't even statistically significant! But then that would mean that L2 cache settings don't matter at all, which seems
counterintuitive, doesn't it?
Mark Kundinger
Hello Mark,
I don't really think 3DMark 99 Max is a particularly consistent
benchmark. But it's useful as an alternative benchmark which we can all
use to verify results from other benchmarks. To minimize the variance, I
have already benchmarked each cache latency using 3DMark 99 (and other
benchmarks too) three times and averaged the results before using them
in this article. So, the variance would be considerably less than the 1.7%
you mentioned.
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
Comment #10
Hello!
I liked the article; what it reminded me of was the bad old days where
one had to worry about HD interleave (pre386). I have an Athlon 500 with
SD11 with no BIOS adjustment for L2 cache latency. I'm looking into overclocking cards and will see if I can get one that will allow me to
set this parameter. Thanks again for writing this article!
Fred Bosick
No problem at all, Fred! :)
Adrian Wong
Adrian's Rojak Pot
http://www.rojakpot.com/
http://www.adriansrojakpot.com/
|
