School the RW in PhD level memory tweaking plz...

Hi guys,

I just lately realised how ignorant I am when it comes to memory tweaking. I have somehow fuddled along using the "back off the timings until it works" method of RAM tweaking, but I have just discovered much to my amazement, that this doesn't necessarily get you the highest FSB, the best stability or good bandwidth. I just saw a system transformed from a finnicky glitchy "meh, stable enough" mediocre overclock, into a smooth, stable, beast by tightening some of the timings. It's like the difference between using an overbored, overboosted 4 banger and a big V8.

So having seen this, I want it all, ALL, muhuhahaaaaa ...... ALLLLLL *ahem* yes, stability, bandwidth and high FSB, my modus operandi was "pick one out of three" because that's all that seemed possible, I'd see peoples benchies and think they got lucky with the stick and mobo or something.

There seems to be too many settings for a simple methodical plod through every combination, and I am beginning to realise that that method can have traps. Like say at 133FSB one set of settings A might look crappy, bad bandwidth, while another B looks better. Then when you pump up the FSB, at 150, A gets the same bandwidth as B, but B is looking unstable, then at higher FSB, A really takes off and B won't even POST. So, following the stickies and the general mem guides out there, I might end up with 150fsb stable, I'm using the settings that are "sensible", maybe I get lucky.

However, I'm getting the idea that there are advanced techniques for REALLY getting your RAM zeroed in. Are there ways to look at a memtest report and not see just PASS or FAIL, but see which parameter might be the culprit for a fail or error on one test but not on others? Can one see the difference between a heat/voltage error and a configuration faux pas? Are there tools beyond memtest86?

I know you've got to put the time in on careful testing still. I just want more insight into the art or science of homing in on the "awesome" settings instead of just the okay ones. Dang, I've put days in on some boards, tweaking and tweaking for another Mhz, then I'll stumble across someone's "optimised settings" on the net, and try them and all of a sudden I've got 10-20-30mhz more headroom, and I'll be thinking, ummm I probably tried that some time ago, why didn't I realise it would clock like that THEN.

So please school me if you can, I want the postgrad course


Road Warrior

Mate,

As with a PhD, the focus should be on something specific, in your case, this translates to a specific setup . General guidelines are useful only upto a point. Spit out your rig and recieve your PhD

First off you need to know what chips you are working with to know your RAM limits. Not knowing this can bring you world of pain;

Good EX... Let’s say you have standard TCCD chips and you go browsing through net and see some guy’s awesome results with really tight timings that get him very nice bench results and you decide to use them yourself.

Well this is where you can make your biggest mistake and most likely kill your RAM, cause not all chips respond same to certain timings and especially Vdimm (RAM voltage). So therefore identifying your chips is the first priority without a question.

Right now in DDR game Samsung TCCD/TCC5 and Winbond UTT/BH chips are the best known players. Then there are Micron which are in my opinion most unique chips out there since they can respond well to either loose or tight timings and with low and high Vdimm. Hynix are another decent chips that are similar to TCCD's but almost always run or very loose timings and somewhat low Vdimm. Infenion are decent and are making their name known on 1GB modules, but are getting mixed results. As far as Mosel chips go they are pretty much dead, but they still use it on value DDR400 sticks.

Big difference between them is that Samsung chips generally like looser timings (2.5-3-3-7 to 2.5-4-4-8) with a lower Vdimm (2.6-2.9V, TCC5's like slightly higher then TCCD) and do anywhere from 200MHz all the way up to 300+MHz.

On the other hand Winbond chips are well known for hitting high freq (from 200MHz and up to 270MHz) with tight timings (2-2-2-5 and in some extreme cases 1.5-2-2-1 to 5), but to do this they require extremely high Vdimm of about 3.6V+. If you were to feed that much too standard TCCD or any other chips in general (maybe except Micron) it would damage them instantly and most likely kill them.


After you figure that out then it's time to test out your RAM and the best tool for that is the Memtest because it will stress your RAM on all levels through 10 different tests.

For more info on that read this and if you have any question about it just let me know.

Also burnning in your RAM is important as well as with CPU's or othwerise you will run out of room fast and will be forced to use high Vdimm to run on a higher freq. For this you can again use Memtest and run it in cycle meaning that you run all 10 tests right after another for a certain period of time (longer is better) and you can do SuperPI 32MB and Prime95 which will also put hurting on your RAM.


That's all that I can think of right now but if you run into a wall don't be afraid to ask for help

Great info RedDragonXXX thanks.

I'm slowly getting my head round the fact that half my problems might be the wrong RAM in the wrong rigs I just had a bit of a revelation, I was looking for something to confirm my RAM specs and found spd_tool on simtel and realised I could translate the various cycle time specs in ns from the spd into T clock ticks for given frequencies, which was a big help. I was looking for such generality, so that I can maybe take home a new chip no-one has had yet, or not many ppl have, and know what to do with it, or what not to do with it. My last big upgrade was such a dissappointment because the TCB3 was just out at the time and there was a heck of a lot of BS around and I didn't know how to sort it out. So that and PSU issues meant I didn't have a cutting edge rig for my money, because it took me months to get it overclocked and stable. So am looking to know more, so that when next I get a big upgrade I can make the most out of it in a week or two and have something a bit special for a few months.

Thanks for drumming it in that all modules are not the same, I'm beginning to see it, I thought it was just voltage tolerance, doh. Re-re-re-reading your and other stickies I think I am beginning to pick out my mistakes. For instance I might have had tRAS too low quite often when I upped tRP and tRCD. Also may have missed the importance of command rate for high fsb and multi module stability. I think I read some guide somewhere that just said "leave it at 1" or other such unhelpful drivel.

What I'm actually messing with at the moment is an AMD 761 board with a 256Mb PC3200U module of Samsung TCCC, now oddly the settings that moved me on, are 2.0,2,2,6 2T command, which is not what you'd expect on this RAM, however, this chipset is reputed not to like cas 2.5, which was probably holding me up for a long while since I tried to be "safe" on that and got nowhere. It seems to be topped at 162Mhz at the moment. Of course could be RAM or chipset at this point, or any number of other things. I see TCCC isn't meant to do such tight timings, it's well under spec though and at 2.8V. Anyway, I'm wondering if 2.0,3,3,7 will do me any good or not for getting it up a few more Mhz.

I might be better off swapping it with another rig though, i've got TCB3 in a KT-333 board, that seemed to run it cas 2 fine async at something like 210fsb, but running sync it only seems to like cas 2.5 and that rig won't clock much without flaking. So I'm thinking i might be better swapping the TCB3 into this rig, and moving the TCCC to the KT-333

mmmhmm, heh, I feel silly in that most of this sort of stuff is probably covered in many posts here, but somehow in focussing to write posts about it, it all becomes clearer.

thanks again,

Road Warrior

NP

My advice is to pick up a mobo, a 7x magnifier glasses , and a miner's flashlight (wear both magnifier and flashlight), oh, you also need a narrow exacto knife. brew a 16 cups of coffee, say god night to spouse, and sit down for a long night, tracing the mobo elec maze. pick where a line starts and ends for about 700-1000 traces. When all down, your first semester of PHD is complete.
No but seriously, I concur w/ super nade, that you have to find out w/ your own setup. I have experienced that highest FSB, even if it passes all tests, is not always best condition and optimal performance. I’ve noticed this when opening some programs.
There are some hidden flaws that none of benchmarks or stability tests detect.