Insane idea for more performance

[negev]

I've just had a completely crazy idea.

Suppose that memory timings 7-7-7-18 are not stable with your sticks, and that 7-7-7-20 is stable.

Suppose you write a program like Memset, that alternated the timings between 7-7-7-18 and 7-7-7-20 every 1-2 (or longer) seconds.

You now test the system and find that the alternating timings are 24-hour prime stable and IBT stable.

Have you gained performance?

What if for every 10-second period you set 7-7-7-18 for 2 seconds and 7-7-7-20 for 8 seconds. Probability wise this would be more likely pass prime. Would this result in a gain in performance?

I'm tempted to knock this program up just to see if my theory works..


Maybe the same might work for fsb/cpu. Imagine if you could have an fsb "pattern" set in the bios, rather than one fixed setting. Something like 421 -> 422 -> 423 -> 424 -> 425 -> 420 (repeat)

Bloodrage "Lunatic Edition" anyone? ;)

[Xello]

You haven't solved the problem of the instability, for half of the time your memory would still be on an unstable setting so you would still get crashes. No? :shrug:

[negev]

Perhaps, but I believe in quantum psychology.

Let's say the ram at 7-7-7-18 causes rounding errors in prime after 2-3 hours. Does that necessarily mean that it will also fail if it's alternating between that and a stable setting?

Consider this - lets say 400fsb is rock solid stable, and 425 causes errors after roughly 4 hours in prime.

So to cause errors at 425fsb you have to prime the system for 4 hours. Now if you program the bios to switch between 400 and 425 at a ratio of 1:6, mathemetically this should in theory push the failure time for the system up to 24 hours. If 24-hour runs using this crazy alternating setting run without error, then you've gained performance.

I'm not saying it will definitely work, but its an interesting idea. Just because something fails after four hours on one setting doesn't mean it can't run at that setting for periodic brief moments of time, while having the majority of the time the drop back down to stable settings and "recover".

[HDCHOPPER]

ask KET he's a bios / ram spd wizard

[negev]

Consider this..

Lets say again that 400fsb is stable, 425 is not. 425 always fails prime, but it NEVER fails in less than one hour. Say you run 50 tests, and in every case it fails after more than one hour.

Why can't you then set the system to run at 425 for one hour, then back to 400 for half an hour, and repeat? Maybe it will crash during the second hour that its at 425... but who knows? Nobody has ever tried this..

[negev]

thanks chopper, i will give him a shout. f*cking cool handle, i love ket ;)

[HDCHOPPER]

Consider this..

Lets say again that 400fsb is stable, 425 is not. 425 always fails prime, but it NEVER fails in less than one hour. Say you run 50 tests, and in every case it fails after more than one hour.

Why can't you then set the system to run at 425 for one hour, then back to 400 for half an hour, and repeat? Maybe it will crash during the second hour that its at 425... but who knows? Nobody has ever tried this..


hahahahaahaa sounds like ya want a workaround for slight instablity problum

altho my old truck with really BAD front end on it if eye whip the streering wheel left & right fairly fast it's like a new truck even at 100 mph !
which without doing eye would never run it past 50 mph :D

eye think your idea would work but at a much faster time rate like miliseconds back & forth

[negev]

yeah milliseconds at a ratio, i.e. 1:50 would be more stable than 1:20. its just a case of finding the sweet spot :)

Saaya get the foxconn engineers working on this :D

[Sailindawg]

Negev, you need to take break from oc'ing this board, however, I do understand your plight with your RAM.:yepp: I'm having a very similar situation with my gear.

In my situation, I chinced and bought some cheap OCZ 1333 RAM. Now, the stock timings at 1333 are 9-9-9-27. I took a chance that I might be able to run them slightly tighter or they might have some head room. Well, leaving them 1:1, I can get them to 1453 GHz (363 FSB cpu) with 7-7-7-21 timings. Using the 333 strap, 266 strap and 400 strap, same thing. Nice, right?! Not quite. The sticks will run 23-25 passes of Memtest v2.01 and always have an 1 error in test 4. Always. At 1.35 vNB, I get errors in test 4, 5 & 6. Up the voltage to 1.4 vNB. Run just test 4 by itself for 10 times, no problems. Run the full suite of Memtest, ONE error in test 4 over 23-25 runs.

Play with the NB GTL, same thing. :shrug: Up the RAM voltage - same results. At lower vNB, when COD4 multiplayer is run, PunkBuster complains that it sees memory corrupted files and kicks you off a server. Up vNB and COD4 MP runs fine, PunkBuster has no complaints, BUT, YOU know that the setting fails test 4 in Memtest.........how soon until either the game is corrupted or I get a BSOD because of the memory.

My solution.......use the 1.3.2 divider with 7-6-6-18 PL6 timings. Runs Memtest fine over night without error in any test. Seems I'll stay on this divider until I get new RAM.

Oh....and the best part about the above, at the higher strap & FSB, where I got one error / 23-25 runs memtest, 12 hours P95 stable, 20 runs ITB stable. Go figure! :shrug:

As for now, I have a decent 24/7 overclock that out performs my old rig. I'm starting to get a handle on my value RAM. What will I do next? Push FSB up to the RAM default speed and start all over again!!!!! :D

In the meantime, I'm going to play some COD4 MP and take break and enjoy my efforts!

Good Luck!!

[Xello]

Well negev the way i see it, the problem here is the way programs like Prime that test for stability work. If a system fails prime at all, no matter if it fails it after 30 mins or after 35 hours, it means that there is instability somewhere and that you will possibly get a crash at some point.

If 425fsb fails after 4 hours that does not mean that you have a set time limit of 4 hours that you can run that before it crashes, Prime isn't an indicator of how long you can go without a crash but more basically whether or not you have instability.

That's been my experience, anyway.. As i understand how the instability occurs through memory etc if there is instability there it has potential to cause crashes no matter how long it takes X program to detect it.

But TBH it comes down to the user what you are willing to accept, personally i want 100% rock solid stable for my system with no chance of a crash whatsoever (does such a thing exist? [i have reasons for this other than purely wanting to avoid crashes/data loss]), but tons of folk are quite fine with 90% stability, who cares about a BSOD every now and then just a restart right?

[negev]

I agree with your views of stability Xello, i want my system rock solid as well.

But let me ask you this - what if you could run 400fsb rock solid but 425fsb always failed after >1 hour. You used a bios hack to fluctuate the fsb from 400 to 425 at a ratio of 1:5 or whatever, and found that you could pass prime for over 24 hours every time. You ran the prime test 5 times and it never failed. IBT passes, super pi passes..... are you happy with this? is this "stable" or not? 1/5th of the time your system is at unstable settings but you can't actually make it crash...

[HDCHOPPER]

the whole purpose of stability is all about data coruption isnt it ?

[negev]

Uhhmm.. what? it's not necessarily data corruption, it's about errors in the hardware. For cpus thats incorrect calculations, retrieval of data from registers being incorrect, for memory its bits being read wrongly.. etc

[Xello]

the whole purpose of stability is all about data coruption isnt it ?

Welllllll i guess so but i mean it's surely possible to have some instability that will cause you BSOD but won't corrupt anything i think? Doesn't really matter i guess...


But let me ask you this - what if you could run 400fsb rock solid but 425fsb always failed after >1 hour. You used a bios hack to fluctuate the fsb from 400 to 425 at a ratio of 1:5 or whatever, and found that you could pass prime for over 24 hours every time. You ran the prime test 5 times and it never failed. IBT passes, super pi passes..... are you happy with this? is this "stable" or not? 1/5th of the time your system is at unstable settings but you can't actually make it crash...

If 425fsb was unstable, it would crash though :D Eventually...

[negev]

If 425fsb was unstable, it would crash though :D Eventually...

Are you 100% sure about that? Nobody has ever experimented like this before... we know it would crash if it was at that all the time, but what if the frequency was fluctuated at the millisecond level? I don't know enough about semiconductors to give any further insight...

What exactly happens inside the silicon if an overclock isn't stable? At a very basic level it would come down to transistors not activating when they should I guess, or vice versa.. the logic is made up of (lots and lots of) transistors. Maybe 1000/second open/close operations is stressful to a transistor.

I still think this is worth exploring....

[Chris_redfield]

Its a great idea but in the end I think it is appealing to the law of averages.

I have had OC's in the past go for hours before failing prime. The trying the same settings again with the same equipment has had prime fail in a few minutes.

While alternating between the two settings may reduce the liklihood of a crash, in the end if you run it long enough you will still run into a stability issue. And sod's law dictates that it's just before you save at the end of the level as well.

[Gendo]

Hmmm the best solution I can think of to solve instability is more voltage:yepp:

[TheGanG]

I would love to have "minimal" VCPU/VNB steps, for example: 1.70V-1.70125-1.7025V-1.70375-1.705V...
(not to have a 1 mile long VCPU-list but i would implement it where i need it)

I believe it serves more stability and OC ability. BIOS modding is a good method to find someone's own values...

[Sailindawg]

Errors in hardware = eventual data corruption. You would need your software to somehow "sense" when data corruption is imminent, then switch to the more "stable" state.


What exactly happens inside the silicon if an overclock isn't stable?

In my pure layman's understanding, as the cpu/memory run further out of spec, in an overclocked state, the electrical signal from cpu to memory degrades. In a very unstable machine, the signal degrades to the point where it can no longer be processed and you get a BSOD.

It's best to think of the pc signal as a wave form on an oscilloscope. Run your rig stable and the wave form is symmetrical and even. It's predictable. Increase FSB, the wave form changes accordingly. Raise the FSB too much and the wave form is out of synchronization, therefore BSOD. In our case, we have a lot of BIOS settings to smooth out the wave form. GTL, voltage.

What would help us out in overclocking is an oscilloscope. This would give us the data set that you're looking for. One would be able to tune your rig to the limits of your BIOS settings very accurately. Unfortunately, we don't have oscilloscopes and they aren't part of your motherboard's package. We are stuck with good old trial and error. The basis for the scientific method.

Here's a good read (http://www.edgeofstability.com/articles/dfi_p35/gtl/gtl1.html) on GTL's, and wave forms are used to explain GTL and it's effect on cpu's.

[Ket]

I can see the theory, but no, it wouldn't work that well and you wouldn't gain much performance, if any. Certainly nothing noticable at least. Primary memory timings are tied to their sub-timings, adjust the sub-timings accordingly and you can at least increase the chances of running the primary memory timings at the settings you want.

[saaya]

yeah, its not like your hardware can run a higher speed stable for some time, but then runs out of stamina.
if prime fails after 4 hours, it wont always fail after 4 hours... it might as well fail after 1 second...
with your idea you stretch the thing... and yes, a crash or error becomes less likely... so if you dont want full stability, this would be a way to get slightly better performance at the loss of a little bit of stability... i guess you can fine tune better this way.
but im not sure its really worth it... when it comes to benching, then every bit helps... ut for 24/7... do you really see a difference between 777-20 and 777-18 timings when browsing your pc, the internet, or playing games? i dont think so... :D
see what i mean?

Essentially i had the same idea when i worked for cellshock, i really got annoyed that there are just round numbers for every timing :D
how come i cant set Cas 4.3? :D

initially there used to be cas2.5 which was really nice!
it allowed some mem to clock as high as cas3 with actually almost the same perf as cas2, notably faster than cas3...
Well, not much we can do about that however... youd have to bark up another tree for this particular one tho, samsung or micron etc...

You might be able to run higher speeds for quite some time without any crashes, depending on how you finetune it, but it will never be really stable.
OC Panel will allow you to adjust the PLL, voltages and other settings directly, and you can save and load profiles and jump from one setting to another profile with the click of a button... takes a fraction of a second... so then you can lower your clocks for demanding apps or games, and run higher clocks in less demanding apps and games :D

[dinos22]

there is no way in hell you will see any difference between the suggested timings either :D hehehe