If you mean to "start all over again," I'd say try 3 Ghz, 1333 FSB, 667 DDR at either "Auto," or the fixed voltage that BIOS voltage monitor reports as auto.
But you say you're stable at 3.348 Ghz and 9x372.
HOW - E-VER!! Looking at your voltages, those settings seem much too high. I think you should get to 3.33 Ghz with the Vcore closer to 1.42V. Read the BIOS monitor to see what the board is really reporting as the NB Vcore and FSB-VTT, because those settings also seem too high. And as Russian said, 2.0V for the memory at this speed seems low.
If you want to run the memory at a 1:1 divider at 372, you may have to drop the CPU multiplier to 8 and work from there. But that's not really an optimal setting in the opinion of quite a few here, although it's an option for getting the memory bus-speed "up there."
If I were to start this all over again, with either the Blitz or my current 680i Striker, I'd start with the memory at 2.1 (unless THAT's the warranty-recommended, but I'm guessing 2.2). I might actually start with VCore at 1.4V. I don't see that 1.3 or 1.35 VHT, 1.35 or 1.40V for NB and VHT are that excessive, but you would like to have settings that are only over-volted to produce results. Testing the effects of these other voltages when you run up against stress-failure after an hour or so is tedious, but probably the only way to do it -- one at a time -- keep notes.
You can also run on a divider <> 1:1, but the 1:1 approach tells you a lot. Really, your best over-clock is probably likely at multiplier 9, and before you start testing different divider ratios and memory speeds that do more justice to the memory's advertised spec, you might want to find a stable OC at 1:1, and then start tweaking the latencies downward.
You can get a lot of bandwidth improvement with lower latencies and good memory. Just boosting the FSB is a crude way to over-clock -- easier, but doesn't give you a full idea of what the system will do at more modest settings. Even so, it's a good idea to start with the memory's SPD latency specs, because you know at the lower speed they won't fail.
If, after you found a good over-clock at voltages not too far from the retail "maximum," you can try tightening tCL to 4 and then even to 3. tRCD is a good candidate: again, from 5, to 4, to possibly 3. If you can get there, tRP can go down to 4 and maybe 3, and tRAS can be tightened to some integer between 6 and 9. Once you have reduced basic latency settings that are rock stable, you can try setting the command rate to 1T, but whether 1T works at that CPU speed or you have to go to 2T, You can most certainly-- absolutely -- positively drop the "advanced" timing tRC down the tRP + tRAS, or that sum plus 1. This latter tweak gives enormous bandwidth improvement.
On the 680i board, with Crucial Micron DDR2-1000's (and 800's would've worked just as well), I think I had my E6600 at 3.3 Ghz, CPU_FSB 370 and DDR = 740 -- FSB=1480. I had the latencies tightened to 4,3,4,8. I was only running the voltage then at around 2.125V and skittish about pushing it higher. But the warranty limit is 2.2V, and there should be no need to be skittish, just "sensible" about incremental increases. And at that time, I hadn't even thought to drop the tRC latency setting from its default (between 21 and 30) to 9.
It would've made a helluva difference, and I wouldn't have felt as much inclined to try different multipliers and higher FSBs, or even change over to the Q6600 quad.
Just remember that CAS or tCL of 3 will probably only work below DDR 730 or 720. tRCD should be more flexible downward. tRP should probably be set = tRCD or tRCD+1, and tRAS should = tCL + tRCD by conventional wisdom. Conventional wisdom based on "DDR1." There is something in DDR2 called "additive latency" which may allow lower tRAS values. And of course, the bank-cycle-time tRC depends on row-precharge and row access strobe values -- with only moderate influence on stability but high influence on bandwidth.
When you start testing latency settings, a tool like Everest Ultimate's Memory and Cache benchtest is good for comparison to see how things improve. People swear by Sisoft Sandra, but you're interested in relative changes with a fairly simple presentation of bench results.
. within seconds after the change, it gave an error. (It was at 1.4240 for the first 20 minutes)
Facebook
Twitter