Blameless

See post #9

Earlier today I decided to throw my Team Xtreem PC8000 back in my system, in order to make sure it still worked the way I rembered before I tried to sell them, and to test the new heatspreaders I installed on them.

So, I normally run my system at 450-475MHz FSB, but I've been using 400 for the last few days as my bearing on my usual CPU fan (a scythe slipstream...horrid reliablity at high temps, but that's another story) locked, forcing me to remove it and run passive, at least until I rotate my heatsink and put my Panflo back on. Anyway, my Team has done DDR-1200 in the past (with the stock heatspreaders) without too much trouble, so that is what I indented to test. However, I could not remember wether my clock skews liked to be advance or delay at this particular ratio, so I set my memory at the lowest voltage (about 2) volts I know it will boot into memtest86+ at DDR-1200, in order to rapidly test varios clock skews, but...no POST.

To make a long story short, I spent the better part of the next three hours testing an ungodly number of voltage, skew, and tRD settings before realising my memory was running much warmer than it should (even for the 2.3v I was now using). Then I remembered something...the P5Q-D has a bug where the 2:3 ratio with the 400 strap is, in actuality, a 3:5 ratio.

I double checked Memset, Everest, and CPU-Z, everything said 3:5 ratio. In order to confirm, I ran some benches, and the 400FSB benches beat my old 450MHz FSB benches by several hundred MB/s and a few ns in latency. I had indeed been running DDR-1333 the whole time. Even more surprisingly, it was almost stable at this point, with no fan over the memory.

After putting my Spot-cool back over my ram temps are now quite manageable (if still on the warm side), and I'm Priming away with no errors so far. The new heatspreaders (Enzotech forged copper, attached with Artic Alumina thermal adhesive)

It's only single channel for the moment, and I doubt it's entirely stable (will pass prime and memtest86+, haven't done HCI yet), but I'm impressed none the less.

After I re-attach the second heatspreader (thermal adhesive set wrong, had to boil it in 100% acetone to get the damn thing off, but everything still works great) to my (even better clocking) second stick, I'll give dual-channel a shot.

Here I am right now (single chan), with an antec spot-cool over the memory and 2.35v real vdimm (2.28v BIOS):

Whodie

great work! those are some nice sticks

__________________

Project: That's tha FIRE!

Blameless

Thanks

Best set of the six (four team of various speeds, and two double sided ballistix PC6400) D9GMH kits I've owned.

Ghostleader

Nice speed.

But what´s up with the bandwith and latency, shouldn´t you get higher bandwith and lower latency, 1066 555 15 tRD7 at FSB 400 on a P45 get better result.

Blameless

Bandwidth sucks because I only have one stick in there ATM (single channel).

Latency sucks for the same reason. The P45 is 3-5ns slower, latency wise, if you do single channel. Also, at 400MHz FSB, each step of tRD is a full 2.5ns (17.5n from the tRD alone).

Absolute best latency I can manage on the P45 is a modest 50ns, and that requires tRD 7 at over 450MHz FSB, dual-channel, DDR-1250ish, with tighter sub-timings.

Once I get the second stick in there bandwith will be over 10k (probably ~10.5k read, ~9.5k write, and ~10k copy, at 400FSB) and latency should be a round 56ns.

Even at it's best, P45 latency numbers cannot compete with the P35, or X38/X48.

Also, it's turning out more stable than thought it would. I'm at 3400% in four instances of HCI memtest, no sign of errors yet. Ram is only luke warm with the tri-cool on high, even at these speeds with 2.35v.

18 is # 1

Super Speed!!!
I think the high tRFC may be slowing latency.
Did you guys see this post suggesting high tRFC mat negate electromigration? http://www.ocforums.com/showthread.php?t=588547

__________________

I fold for OCN

http://valid.canardpc.com/show_oc.php?id=643273

XFreeRollerX

Dude your bandwidth is in the hole like you said lol

my 1000mhz 5-5-5-15 RAM does 11,000 GPBS with 4 sticks But GREAT OC man!!!

__________________
Currently playing: Battlefield Bad Company 2 under the gamertag "TheShoura"
My Awesome HTPC/Server/Gamer Machine Computer related images
Benchmarks Social Clubs and MORE!

Blameless

Interesting hypothesis.

Take out three of your sticks then tell me your bandwidth isn't in the hole.

Dual-channel is only a few hours away.

All the old epoxy has almsot disolved off my second stick, and my heatspreaders are cleaned and polished.

Blameless

Alright, been doing further testing and after getting some more strange results I have to conclude that this is like not DDR2-1333, but DDR2-1200.

I went up to the DDR-1280 ratio, which was reported correctly, and got (marginally) higher benchmarks, so my previous conclusion must have been mistaken. Apologies for the misinformation.

Seems that the bug in this BIOS is more complicated than I had previously thought. Most of the 400FSB strap ratios are giving me weird weird problems (they are either reporting contradictory speeds between the BIOS and various programs, not performing as they should, or being far less stable that other, faster ratios), so I think I forego further testing with them.

Also, testing at 400FSB is making results difficult to interpret. The change in performance at ratios higher than 3:4 (where tRD7 becomes an option) is so miniscule that it very difficult to see if anything has really changed.

Now to find that thread on Xtreme Systems concering the DDR1200 ratio in the 400FSB strap actualy being 3:5 instead of 2:3 when it's actually been 2:3 all along, just reported as 3:5 in software and giving some bizzard results.

Man, if only I had an occiliscope this crap would be much less tedious.

Blameless

Looking at the difference between single and dual channel it's more understandible how I could have originally mistaken this for DDR2-1333:



Benches barely change. Ironically, latency sees the largest improvement in going from single to dual channel.

With so high a ratio, single channel is nearly saturating the bandwidth potential. FSB has to go up to see an appreciable boost. This may also illustrate a limitation in Everest's benchmark. I don't think they use a large enough chunk of memory, so they are testing cache performance in the "memory" benchmark.