832MHz 400 strap v. 333 strap <- one works one doesn't

[DMF]

Asus P5Q Pro, 1306 BIOS, E8400, 2x2GB Mushkin DDR2 HP2-6400 5-4-4-12

memory timings set manually ^^^
FSB 416 (CPU 3.76GHz @ 1.26V (1.30V setting))
CPU is not the limiting factor; this happens even when I drop the multiplier

with the NB strap set at 400 memory divider is 1:1; DDR-832
with the NB strap set at 333 memory divider is also 1:1; DDR-833

The 333 strap is preferred because of lower default tRD http://www.anandtech.com/showdoc.aspx?i=3208&p=4

--------
Problem: 1:1 400 is stable; 1:1 333 is not.

1:1 333 sometimes won't POST, requiring a complete power cycle (AC switch off). When it does POST it produces "BOOTMGR is corrupt" when Vista tries to boot.

memory tested at 1.82V, 1.90V, 2.00V
NB tested at Auto and a couple values I don't remember.

I don't get this at all. Suggestions?

 

[n7]

Why are you not setting tRD manually?

There's no need to leave it at auto as your board allows manual adjustment.
(Look under AI Transaction Booster > Common Performance Level)

Other than tRD, other internal NB timings/speeds are usually affected by using the lower vs. higher straps, which may also be contributing to the 400 strap being stable vs. not on the 333.

I would loosen up tRD & see what you can tighten it to for the different straps, & then bench to see what is best.

You might want to bump NB voltage up; that might also help you in your goal.

Really though, TBH, i would not be trying for 1:1 operation; i'd use it as a last resort.

On the P5Qs, 1:1 requires a far looser tRD than with 5:6 or 4:5, not to mention you gain bandwidth from the higher speeds as well.

You said FSB @ 416, so with 333 strap 5:6, you'd have DDR2-998, which your RAM should be able to do.
Or 266 strap 4:5 (this will be harder to do, as it's harder on mobo & higher RAM speed of course) would be DDR2-1040.

 

[DMF]

That's what I was doing when I ran into this issue. The article suggests that I should be able to use tRD=8 as a minimum (8-5/1 > 2/1). When I set 8 from Memset it locked up the system immediately. (I gather NB needs more voltage.) I can run 10 (Auto vNB with vCPU 1.325V (setting)); haven't tried 9.

Since I haven't been able to find a program that shows vNB, I'm assuming that Auto bumps it up just as Auto bumps the vCPU. So I think it's at 1.30V already. What do you suggest I try?

I'd like to try pushing the CPU higher, though current 3.76GHz may prove to be the limit on stock cooling. I've had it running at 432 with voltages on Auto. 1:1 gives me the headroom I need. I'm thinking that 1000 is going to be the limit for this memory. I tried setting 6:5 998 @ 2.0V and no go. 2.1V is max. I'd prefer not to run it that high for the final settings, so 6:5 is only a marginal option.

 

[DMF]

More experimenting. I cranked the NB to 1.40V, RAM to 2.06V (mindful that Asus boards are reputed to over-volt by 0.1V) and was able to run 333 strap 6:5 DDR-962MHz @ tRD=10, though not stably. I tried tRD=9 (Memset) but it locked up. Obviously I was also able to run 333 strap 1:1 at those voltages.

Meanwhile I can run 400 strap 1:1 DDR-832 @ tRD=10 at vNB=1.24V (or less maybe) and vRAM=1.90V (or less). I had to bump vCPU to 1.350V (setting) because Prime95 failed last night. Right now I'v started another Prime95 run (Blend) with the manually-set voltages.

Some questions:

What are the odds of being able to use my CAS-5 sticks with a CAS-4 setting? That would(?) make a big difference in access latency.

Why does a 1:1 divider display 3 phases in the Ai Transaction Booster section?

Is there a decent free tool for showing actual access latency times and/or evaluating the performance of the current memory settings?

TIA

 

[n7]

The X48 rules do not apply to P45.

The concept still does, but the P5Qs do not run the same tRD values as the board in that review, nevermind other boards.

The value you can use does vary board to board.

Don't use auto voltages; auto overvolts a bunch of them way more than necessary.

See my post in this thread on the same thing: http://forums.anandtech.com/me...id=28&threadid=2249081

You need to be testing for max CPU & max RAM separately.

IOW, for the CPU, loosen tRD to say 12, RAM @ 1:1 5-5-5-15, & then find the CPU's max based on whatever you are basing your max on (vcore/temps/speed, etc).

Once you know that, then work on the RAM separately.

IOW, set your multi very low, & use a higher ratio, to keep CPU out of the equation.
Set tRFC very loose - try 60+, keep tRD loose - keep it at 12, & test whatever timings you want.
Find out max 4-4-3, max 4-4-4, max 5-4-4, max 5-5-4, max 5-5-5, etc.

So say try 7x300 3:4 (DDR2-800) & Memtest86+ (particularly test 5) + HCI Memtest as you increase the FSB till the RAM runs outta headroom.

Now that you know the CPU's max as well as the RAM's, work on combining those results for the best performance.

Trying to figure out maxes for both at the same time is not going to end well, as you'll have a hard time figuring out what is wrong.

Always prioritize CPU speed, & work tRD (which will be dependent on FSB speed & ratio) & RAM speed around that.

Use Everest for showing differences in RAM latency, bandwidth, etc.

 

[DMF]

I've made separate passes at CPU and DRAM limits. CPU is limited by heat (~60°C Prime) at around 3.85GHz, but that is of course dependent on voltage, which I've been reading up on today both in that thread and the sticky Guide. I can probably bring the vCPU (1.350V setting) down by tuning FSB VTT and vNB. I found the DRAM limit for 1.8V 400 strap somewhere above the 832 I'm trying to run now (847?). But I need to do it again with a better understanding of how the various bits relate. Thanks for the suggestions.

First I'm going to change the Auto voltages. Then I'm going to see what I can get away with at CAS-4.

as you increase the FSB till the RAM runs outta headroom.

You don't mention vDRAM. Think I should just set max (2.1V) and leave it there for this sequence?

 

[DMF]

Okay, at 2.1V & timings much as you suggest, CAS 4 doesn't even make DDR-800. <boo> 5-5-5-18 is solid (so far) at 951, as has been reported by others with the same sticks. 5-4-4-12 (advertised timings) POSTs at the same speed and will likely clear Memtest too.

So the RAM is quite capable of what I'm asking it to do (832). If I cap the CPU speed at 3.8GHz, I wonder if a lower divider with a faster FSB and lower CPU multiplier will let me run, say, DDR-900...