Originally posted by: Ararat
I think I got the cultural reference this time! I looked that place (which I'd never heard of) up on wikipedia, and I assume you're refering to the racing there.
In regards to the ICH9R, I'm not dissing it. Up until that one incident, it has been reliable with my RAID arrays, and as soon as I switched SMART back off, everything was back to normal.
On to the important stuff...
Throwing caution to the wind, I just jumped straight to 3.3GHz (366x9, 1.45v vCore, FSB voltage +0.2v, northbridge voltage +0.1v). I ran 4 threads of Prime95 for 5 hours 10 minutes with no issue. Temps were ok (62-65 in Intel TAT).
However after some use that day (H.264 encoding, downloading etc), I got a BSOD with a 'driver irql not less or equal' in reference to tcpip.sys. Now that sounds like a network driver problem, but it's a clean install, and everything's up to date, so I couldn't be sure it wasn't oc related, so I throttled down to 3.15 (350x9), and am leaving it there for a while. If it happens again, I'll know it's not oc related, and proceed to squeeze as much juice out of it as possible (though I'll have something new to pull my hair out for).
The fact the Prime95 went for 5 hours with no issue suggests to me that the BSOD isn't related to the overclock, what do you guys think?
You guys seem to be talking a lot about memory dividers. Why? My strategy tends to be to just run it in synch (1:1, or as the motherboard calls it, a multiplier of 2), which means memory shouldn't be an issue unless my FSB passes 533. Once I settle on an overclock, I'll just pick the highest multiplier that doesn't result in overclocked RAM. Is my strategy flawed? (Admittedly, I have far less experience with these things than most ppl on these forums).
I don't think you have a flawed strategy, but on the matter of over-clocked RAM, consider as follows.
800Mhz and 1000Mhz DDR speeds exceed specs on most processors I know of. And these DDR2 standards arrived on the scene a couple years ago. So it seems a lot of motherboards default to a "native mode" of DDR2-800 if indeed the SPD information shows that to be the base speed or higher.
I discovered that with 4:5 ratio and DDR2-800 @ 875 Mhz and timings lower than the DDR2-800 spec, I could run the memories at 0.025V lower and actually get a just slightly-higher result with a synthetic bandwidth test -- when compared to much tighter latencies @ 1:1 and DDR2-720. In other words, 360 was as high as I could take the processor without increasing the voltage again.
I look carefully at RAM kits before purchasing them, to see whether latencies can be stably lower at lower speeds. In some cases, it wouldn't make any difference purchasing DDR2-800 versus DDR2-1000 (or 1066) modules -- it would be possible to have "flexibility" with the 800's if they over-clock to 850 and above, and I can have flexibility with DDR2-1000's if I can tighten the latencies and run them at anywhere between 800 and their rated speed (at looser latencies). There is a three-way trade-off between votlage, clock-cycle-reduction, and memory-bus speed.
Also, I've never taken time to validate this before, but the transfer of data between CPU and memory are not the only aspects of memory performance. You would think with DMA and some other operations, there might be benefits to memory-bus speed even for dividers <> 1:1.
But it's mostly a question of flexibility, and I would prefer a 1:1 ratio in most cases.
And I still believe your FSB limitations with a Q6600 may leave you below 400 Mhz, and in fact, at 360 with the B3 stepping and possibly 380 with the G0. Whether you're better with a 1:1 versus a 4:5 divider under those circumstances, I leave it to you. And -- yes -- you MIGHT get to 3.6 Ghz, but I think it's pretty "iffy" on air-cooling. You move toward a lot of probability-distribution "tails" that way.
On the testing, maybe you know already -- what I do is look for a quick failure under PRIME95 -- for a combination of VCORE, VDIMM, external-clock. I overvolt the memory slightly for certain speed ranges to eliminate RAM as much as possible as a possible cause of instability.
For instance, on this E8600 I'm running, I found a point at 4.1 Ghz where a core failed Blend Test in 17 minutes. When I pushed voltage up just one or two notches, it ran for hours -- 16 hours to be precise, for either Small FFTs or Blend -- no errs or warns. I then kept the voltage settings for 4.1 Ghz, and dropped the external clock by 10 Mhz or 4.0 Ghz. I can run IntelBurnTest on that configuration all day long if I want, but it isn't practical. I ran 30 iterations twice -- once with the latency settings that proved under 4.1 Ghz, and again with tighter latencies.
If you're going to over-clock that puppy for 24/7/365 operation, you need to certify PRIME95 operation for longer than 5 hours (people would say a day), or verify with a few hours of IntelBurnTest.
EDIT: On the BSOD thing, since the screen reported this may be a problem with your LAN/ethernet hardware, reinstall the drivers. Have you installed the chipset drivers? Don't use Microsoft drivers when there are updates for same at the manufacturer's web-site. When XP or VISTA asks me to do "Express (Recommended)" or "Custom" updates, I pick the latter, and I keep all hardware updates through Microsoft out of the equation.
It may be that the BSOD is an indication of instability with your over-clock -- and only randomly linked to the LAN drivers.
Facebook
Twitter