a64 and memory speed - the truth

C

(cf)Eclipse

Freelance Overclocker
Joined
Feb 18, 2003
Messages
30,027
ok.. since i'm starting to get frustrated with all the people who believe that they must have awesome ram to overclock and have the results be worth it, i did some quick testing. yes, i know it's not very thorough.. i'll probably attend to that another time when i really have some spare time on my hands, not just as a quick saturday project. but anyway, on with the results.

here's my system as used for testing:
a64 3200+ clawhammer - clocked at either 220x10 or 245x9 for 2.2ghz
edit: also added 240x10 for 2.4ghz
2x512mb crucial ballistix
geforce 6800gt at 375/1100
w.d. 80gb ide w/ 8mb cache - application drive
w.d. 250 sata w/ 8mb cache - data drive
ocz powerstream 420

since i originally limited myself to 2.2ghz on the cpu for the first few tests, i was left with three ram speed choices:
220mhz (1:1 using 220x10) @ 2.5-2-2-8
200mhz (5:6 using 245x9) @ 2-2-2-8
183mhz (5:6 using 220x10) @ 2-2-2-8
as an update to this to show how cpu speed really is more important than ram speed, i also ran the tests at 240x10 with the ram at 5:6 with timings of 2-2-2-8.

i used 8 because with quick testing using a64tweaker, tras of 8 gave best bandwidth/latency on average with my ballistix
if you don't know exactly how a64's memory controller sets ram speed, have a look at the link in my sig.. it'll also explain why the 200mhz option is really 200mhz not 204mhz like you would think at first

ok, so first things first, the semi-synthetic benchmarks. i'll start out with the results from 3dmark05. i know the final score is very gpu limited, but it also gives a cpu score, which has more releveance to what i'm trying to prove here.
3dmark05.PNG

see how the overall score changes.. none. disredarding that random spike at 200mhz, which is within the margin of error for such a benchmark. it shows the cpu score scaling with ram speed though. not much, but a little.
now look at how much extra performance a 200mhz boost on the cpu gives. 9% core speed gives 23% cpu score improvement. something strange is up there, but i ran the test again, and it gave me the same result, so i don't know what gives. i'll go through all of them again and see if i messed something up.

next is (arguably) a bit more relevant. everest read/write and latency scores. nothing unexpected here, nor anything that needs to be explained, other than nore how going from 200mhz @ cas2 to 220mhz @ cas2.5 results in the same latency.
memory-bw.PNG

memory-latency.PNG

(note that these results are the average of 3 tests with correct sig dig's ;))
one interesting thing to point out. notice the decreased latency by increasing the cpu speed 200mhz... not a lot, but it's there.


now for some more practical tests. i like to make backgrounds using a terrian rendering program called terragen. the test was to make this picture. then i used dbpoweramp to convert the scooter - 24 carat gold album from mp3 to ogg vorbis at a quality of .6 (~192kb/s iirc)
time-based.png

here you can see that memory speed does have a little bit of difference, but nothing profound, especially considering that i have an extra 600mb/s and 4ns lower latency to play with between the fastest and slowest settings.
then look at the extra 200mhz, see the profound difference? i like the scaling with the cpu speed. 9% more speed, 9% less time for both cases (actually 9.6% and 9.4% for the respective tests) :D

and now for real world testing. i have one game that in game benching capabilities: doom3.
*heavily updated*
so i ran two tests. one high (high quality, 1280x1024) and one low (low quality, 640x480)
i recorded the results from the second run to eliminate the errors. i disabled vertical sync in the game to eliminate the monitor limitation. sadly, i don't have the resources [H] does, and couldn't run the demo and get a fps graph, so an average will have to do.
doom3.png

it seems memory bandwithd has a slight effecthere.. however at 200mhz, 2.2ghz, there is an interesting anomaly, though 1fps isn't really anything to get too upset about, it is about 1% after all :D

so here's my conclusion: for gaming, memory bandwidth doesn't seem to be too important, latency might be, but due to limitations of my memory controller, i can't get under about 44ns, which i'm pretty close to already.. i would like to test some other games though, like hl2, but i don't have that game right now, so...
for other things like encoding/rendering and general computer usage, memory bandwidth helps, however very slightly.
for all you oc'ers out there that demand high mhz ram for your system, i can tell you one thing that it makes a notable difference with: benchmarks. in super-pi, going from 183mhz to 220mhz dropped my time from 41s to 39s, and you can see how much it effected my cpu score in 3dmark05.
the only other real reason to have nice ram is to have more flexibility with ram settings, so you can run at weird multi's and dividers to get that optimal cpu speed and not have the ram underclocked or something.

next step: get a 939 system and run benches in single channel and dual channel mode.. see how much 2x the bandwidth really does help :D
 
nice work.

i have a feeling we'll see even less of a difference when we bring in dual channel.
 
thanks, and yeah.. though if i go 939, i'll have to get a chip with 512kb of cache cause i can't afford an fx.. and with 1/2 the cache, the changes in bandwidth may be more apparent.
 
damn this is very interesting. from cas 2 with lower HTT to cas 2.5 with higher HTT the higher HTT loses. Am I reading this right? I might have to go over it a couple more times.

I have never used an A64 yet so I'm still very confused.

( did I use HTT in my sentence the right way? )
 
MaMMa said:
damn this is very interesting. from cas 2 with lower HTT to cas 2.5 with higher HTT the higher HTT loses. Am I reading this right? I might have to go over it a couple more times.
Click to expand...
yeah, that's about right.. and if you think of it from an absolute latency standpoint, the fact that they're about equal makes sense.. just multiply 1000/mhz * latency and add them all up. it's 27.3ns for 220mhz vs 30ns for 200mhz. of course there are other factors that are involved, but it's a good rule of thumb for picking faster latencies.

and yeah, i've seen that article over at anandtech, but that's purely hl2, i want to see things that i do more often, like rendering performance, or how long it takes to convert music from one format to another.

at the moment, it seems like cpu speed is the overwhelming factor in most applications, but in things like gaming, the extra cache on the clawhammer helps out enormously. but then again, some things (like apparently hl2) benefit a lot from the extra bandwidth dual channel provides, and it makes up for the lack of cache. it all depends on what you're doing..
 
i know i shouldn't be bumping, however i made some changes to this. for everyone who has already read it.. please re-read so you can see how awesomely the a64 scales with clock speed :D
 
can you label what each graphs information is? like what test is what and what measurment stuffs in? my eyes are too bad to try and sit there and figure it out :(
 
i'll put labels on the graph soon, however, if you read what i wrote, i say what each graph is testing, or at least i believe i did.
 
I could add something about mem bandwidth with the 3200 Winchester. all tests with 1:1 divider:

10x200 2.5/3/3/7 = 5800
10x255 2.5/3/3/7 = 7000
10x270 3/4/4/10 = 7400

But in games and SuperPi, only the pure high CPU speed, even with the memory divider, make the most of the diffences.
 
Nice reading there :0) it pretty much confirms what some dudes over at AT-Forum has said for a long time. Its better to scale your CPU up and let the Memory run tight latency and ar a ratio 5:6 then loosing the timings up and go 1:1.
 
That Anand Article is very nice. Interesting to see real world differences between Dual and Single channel, 2-3 CAS, and 512 vs 1mb cache.
 
there's actually a new article on anand that indirectly answers a lot of my questions.
pentium-m, however, when you get to the benchmarks, pay close attention to three a64 chips, the 4000+, 3800+ and 3400+, which are all at 2.4ghz
as many of you know:
4000+ = 1mb, dual channel
3800+ = 512kb, dual channel
3400+ = 512kb, single channel

this way, you can tell what things are effected by what. it seems like there are a few programs that really like the extra cache. more like extra bandwidth, though cache still helps, and others are entirely dependant on cpu speed (there is very little difference between the 3400 and 4000 in these)
 
I have a pc with:

3500 winchester
msi neo 4 ultra
2x512 ocz gold vx voltage extreme

and i need help how overclocking in this machine.

Thanks
 
(cf)Eclipse said:
there's actually a new article on anand that indirectly answers a lot of my questions.
pentium-m, however, when you get to the benchmarks, pay close attention to three a64 chips, the 4000+, 3800+ and 3400+, which are all at 2.4ghz
as many of you know:
4000+ = 1mb, dual channel
3800+ = 512kb, dual channel
3400+ = 512kb, single channel

this way, you can tell what things are effected by what. it seems like there are a few programs that really like the extra cache. more like extra bandwidth, though cache still helps, and others are entirely dependant on cpu speed (there is very little difference between the 3400 and 4000 in these)
Click to expand...

Ohhh da...n Ive allready looked thrugh that article. Could someone do a summary on the important parts... :D pretty pls.
 
haha, sorry frallan, i don't have enough spare time to do a review of a review for all you lazy bums out there ;)


colibrisnegros said:
and i need help how overclocking in this machine.
Click to expand...
and my apologies colibris, but this has nothing to do with my thread. if you want help oc'ing your system, do two things:
read through the a64 oc data thread.
go look in the overclocking forum.

if you have trouble finding an answer right away, try the search tool. if that fails, fire off a message to me, i think i can help you out, though i have no experience with the neo4.. you might be one of the first ones here to have one.
 
My understanding has always been that A64's want very tight timings (low latency) for best performance, and P4's want very high bandwidth for best performance.

So with A64's you go as low latency as possible and raise the FSB to what you can get with those tight timings.

With P4's you raise the FSB to as high as you can, and then tighten the timings to what you can get with the higher FSB.

==>Lazn
 
there's two sides to that lazn..

while timings obvious effect latency, there are two other factors:
memory speed
cpu clock speed

the combination of memory timings and speed make up most of the equation, however, as i pointed out in my original write up, an increase of just 200mhz on the core improved latency too.

and, when i have time, i'll also add a few more graphs on variable timings with a fixed speed. any suggestions of what speed? i'm thinking just 200mhz, i'll do 2-2-2-8, 2.5-2-2-8, 2.5-3-3-8 and 3-4-4-8, though the second one probably isn't needed, i'll put it in, cause that's what my ram normally runs at unless i'm under 210mhz or so
 
You must log in or register to reply here.
Back
Top