Hmmm Higher FSB @ Same Clocks Better Results?

[njkid32]

I am going to test this out and see how it works.. I just wasnt sure if it even matters in WCG or not. Anyone seen better results by doing this?

[sierra_bound]

I don't think high FSB makes much of a difference in WCG. I've run a X6800 at 277X13 for months. Still scores well. This enables me to run the RAM (D9) at low clocks and low voltage. I think what some people are saying about D9 is true. Running at more than 2.2v around the clock is asking for trouble.

[[XC] itznfb]

yea i think crunching these work units will benefit most from raw GHz. higher the CPU freq. the more work gets done.

[[XC] Jaco]

I 'm not sure high FSB would have an influence on the benchmark ? I don't think so. But it's possible it's crunching a little faster with a higher FSB.

I would like to hear Meshmesh view on this

[njkid32]

I just ran 417x8 @3.36ghz and it wasnt any higher in the bionc benchmarks than 370x9 @3.33 so I think that solves that for me:D

[[XC] Jaco]

yes , but it's still possible 417x8 will get the work done a little faster , although the benchmark is the same.

[njkid32]

Also, running a high fsb requires more MCH and Vcore volts.. So, if I can run at stock volts at 370x9 I can stay much cooler.. Once my water cooling gets here I bet I can do much more at stock vcore. I think I am having issues with this mem so I am going to throw in my 9136 Dominators to see if thats what's limiting my oc. On a diff set of mem I was able to do 3.4ghz no prob at stock volts and now all of a sudden I cant. Weird:D

[[XC] Kayin]

Only if WCG is really read/write memory intensive would it start to make a difference, but the scores I'm seeing returned from varying machines makes me think it's more FP than anything...

Makes me wonder, is there a GPU client for the 8800 series? I could do some real damage there...

[[XC] moddolicous]

The only folding project I know that really depends on stuff other than the raw cpu speed is D2OL, which favors tighter mem timings. Does someone want to check that also? Like DDR 800 @ 3-3-3-8 vs DDR 1000 4-4-4-8. I think C2d's prefer more bandwidth than tight timings, right?

[[XC] Kayin]

I can experiment with timing on my AM2 here..

However, if you want tight timing tests, my wife's is running 3-3-3-10 and I just set her up on WCG... I'll see what happens. She's running a PD 820 w/2 GB of RAM, it should be interesting...

[[XC] riptide]

Mod... SOB I know likes tight timings :)

[[XC] itznfb]

Makes me wonder, is there a GPU client for the 8800 series?

i wish. if there was a GPU client for WCG we'd be caught up to easynews by now.

[[XC] Kayin]

Well, I have access to lots of PCs, as I repair for practically the whole area...

[Vapor]

In my experience with BOINC benchmark....

1) cache size on any architecture makes a minute difference at best
2) RAM timings mean nothing (3-3-3 vs. 6-6-6, no difference)
3) Going from dual channel to single channel with an X2 or a Conroe is only a 1-3% drop
4) processidletasks on a 'busy' install can do a lot to increase scores
5) Linux 64-bit and OS X 10.4 are kings of the benchmark

[STEvil]

chipset strap may play a role...

Looking into what helps Super_Pi 32M run faster would be the way to go.

[[XC] moddolicous]

riptide;2135975']Mod... SOB I know likes tight timings :)
I ran SoB for a little while and didn't have much time to mess around with it. That is good to know though.

[Vapor]

chipset strap may play a role...

Looking into what helps Super_Pi 32M run faster would be the way to go.32M is almost pure strap setting and high RAM speed....RAM means so, so little for the BOINC benchmark it's not even funny. Strap may have a role here....never tested that, always just used 1066 as 800 clocks like crap and 1333 is as slow as an overencumbered elevator going up.

[STEvil]

The benchmark does not accurately represent work done.

470x9 may not make higher ALU/FPU numbers than 385x11, but as seen with Pi32M (which does not reside in the CPU's L2 cache like 1M can on 4MB C2D) it will reduce the time taken to produce the same amount of work.

[[XC] gomeler]

SPi is a HUGE I/O benchmark, that is all it is. My testing with C2D/965 show that 3-3-3, 4-4-4, 5-5-5, 6-6-6 timings mean nothing in actual crunching performance. I tested in the sieving projects as I could get instant results. Single channel vs dual channel didn't do much if anything, high FSB vs low FSB didn't do anything, and timings didn't do anything. Clockspeed bumps were the only factor that really mattered, I could easily see a 10MHz clockspeed bump resulting in more performance bumps than timings/straps/etc.

[STEvil]

Did you use the same work unit each test? Is seiving using the same calculation algorithms/etc as WCG?

This is why I am a strong proponent of WU-style benchmarks in DC projects.. too bad WCG does not allow us to test these factors :(

[[XC] gomeler]

Sieving runs the same algorithim billions of times to test numbers, it's the same simple FP operation over and over again I believe. Testing these factors in WCG would take weeks as you'd have to factor in quorom results, scoring time, and random weirdness that is WCG. Perhaps we could talk to the devs of WCG and see if we could secure a single WU that we could utilize for benchmarking purposes? Just tell them it'll result it increased output from us, I'm sure that'll perk a few ears?

[[XC] Kayin]

Simply knowing what will help us crunch faster will result in more work done. We're smart and resourceful little asses.

[STEvil]

gomeler - if we take a single WCG work unit that we know takes X amount of time on X system, then run it on Y system (Y being yours, X being a base system) then we can create performance matrix tables.

Allowing the users to run this benchmark by themselves allows then to test things like ram timing and speed.

One single work unit run once. It is not a quorum result and it never changes. Maybe have it be 15 minutes once a week or whenever the user chooses.

[[XC] gomeler]

How would you consistently run this same WU though? Could we just copy a WU and force the machine to run it and see how long it takes to crunch it? I didn't think it was possible to do this with BOINC/WCG? It would be nice though to have a benchmark WU, something that doesn't take 8 hours on a P3 :p:

[Vapor]

You could have it like the D2OL benchmark....just something on the side, nothing required.....just a way to test and tweak :)

[[XC] riptide]

Gomeler... I'm not sure about the difference between the SOB and Sieving algorithms. I do know in SOB the calculation is mostly resident in the L2 cache. However, I can bet my life that timings INSTANTLY affect the rate of calculation in SOB. And can also be seen over the long term aswell.

[[XC] moddolicous]

We can't use the D2OL benchmark because timings did make a difference. Stevil brought up a really good point: If the timings don't affect the benchmarks, do they affect how long it takes to complete a certain WU? This is a good question.

[[XC] riptide]

I also have a question whether or not a program with SSE2 optimizations would benefit more with tighter timings?

[[XC] Jaco]

The benchmark does not accurately represent work done.

470x9 may not make higher ALU/FPU numbers than 385x11, but as seen with Pi32M (which does not reside in the CPU's L2 cache like 1M can on 4MB C2D) it will reduce the time taken to produce the same amount of work.

that is exactly what i'm trying to say.
It's difficult to measure , but your ppd should be slightly higher.

[STEvil]

If we took a short WU (shorter the better probably) and bundled it into a working BOINC directory we might be able to test that way. Just disable the 'net access.