calculating the "default" tRD

[HJaLTe]

Ok, so I was reading Kris Boughton's article about ASUS Rampage Formula (X48) and how he emphasizes the importance of selecing the right straps and dividers based upon the "default tRD" (MCH Read Delay)

I know how to calculate the latency at the different strap/FSB settings - thats pretty basic - but how do I calculate the tRD (MCH Read Delay)?

... How do I find the tRD (MCH Read Delay) for a given DDR rating??

Did I miss anything in the article, where he explains it or is the Default tRD just a "fixed" value???

I apreciate any helpfull ansawer.
//HJaLTe

 

[HJaLTe]

Is there anybody out there that have the slightets idea of what I'm talking about???

 

[rge]

regarding tRD being fixed....tRD is adjustable on most all boards, though it may be labeled differently than tRD. I can use tRD of 7 (tight as possible) up to about FSB of 480 (ddr2 960 1:1), though I have to increase nb voltage slightly as I increase FSB for stability. Past 480, I have to loosen timings, tRD to 8, to be stable. Since all mobo are different, there is nothing to calculate...it is trial and error to find the tightest tRD timings for each FSB setting. Other than that, I could not understand what else you are asking.

 

[n7]

tRD is more dependent on FSB than memory speed.

If you run a lower FSB, you'll be able to run lower tRD.

I haven't tweaked it heavily, but i can run tRD 5 for my system in sig, since i'm only at 350 FSB...when you run higher, you'll be doing 6 or 7, or even 8.

I'd suggest starting at something loose, like 8, & then tightening it up once you find the CPU OC you want.

If your mobo doesn't allow you to manually set tRD, then you can experiment with different RAM ratios, & you'll see that generally, the lower the ratios, as in 1:1 vs. 2:3, etc, the worse the tRD.

If you don't already have it, get MemSet.
http://www.tweakers.fr/memset.html

 

[Syzygies]

The article in question is:

ASUS ROG Rampage Formula: Why we were wrong about the Intel X48

Tony over at the OCZ support forum posted a tRD calculator using the same formula, so I asked about it:

Re: tRD(performance Level) Calculator ver1 DD2 ONLY Intel Chipsets

(Credit means different things to different people. A performer can seem obsessed with getting their name around, but they're worried about being able to eat. An academic truly just wants to know what to read next. Missing credit to us is like a hanging pointer to a C programmer.)

So the story is that Kris Boughton and Tony started the tech repository before Kris came to AnandTech. Kris found this formula through Intel sources he doesn't wish to divulge (I struck out trying to google for this), so at a minimum the formula applies across a chipset, though each board maker gets to fiddle with BIOS as they please. In any case, the formula is close but not exactly correct for me, and it only predicts posting, not stability.

As I say there, the equals sign is in the wrong place for reasoning intuitively about the formula. The thing to do is to convert both tRD and memory timings into nanoseconds.

Edit: What follows is independent of the chipset; I'm ignoring the X48-specific constants in the article. I have a P35 chipset.

In my spreadsheets tracking overclocking experiments, I'll have columns

A: FSB (e.g. 356)
B: DDR (e.g. 1068)
C: tRD (e.g. 5)
D: CAS (e.g. 5)

DDR is computed from FSB and the memory multiplier, but how one says this varies by BIOS.

Now, to compute (for row 2) Trd, the time in nanoseconds that the MCH waits around, derived from the FSB clock, I use

=1000*$C2/$A2 (e.g. 14.04)

To compute (for row 2) Tmem, the time in nanoseconds that CAS takes for the memory sticks, derived from the DDR clock, I use

=2000*$D2/$B2 (e.g. 9.36)

Now, Trd needs to be a longer duration than Tmem, so the MCH has time to finish its part of the work and pass it on. The formula Kris gives quantifies this for some boards. I find it more useful just to look at the two durations, and decide on hunch and experience if that particular overclock has a ghost of a chance of working. The sample numbers have plenty of room, yet one click down doesn't work for me.

A lot of pronouncements get made in this business, of the form "higher FSB is better" and so forth, that would make sense if our choice space was a giant blob, and we were trying to move to the sunny end of the blob. In fact, our choice space for a given cpu speed is a finite set of points, and the best point may not be near the end of the blob that received wisdom asks us to use, for entirely quirky reasons depending on our exact configuration. Changing the FSB by one click can entirely change this picture, rearranging the stable point cloud. This is a famous mental image for engineers, it's the difference between linear programming and integer programming.

In practice, what this means to me is that for a given overclock, I want to make a spreadsheet listing all my options, sort the rows by estimated performance, rule out rows as impractical e.g. too tight a tRD or too high an FSB, then start stability and performance trials to see what works.

I've been regressing various computations that are realistic for what I want to do. Some exhibit excellent cache behavour, and memory settings basically don't matter. Others thrash the cache, and memory is still minor compared to cpu speed. In these cases, tRD and CAS are about equally important for me, so, yes, tRD is definitely worth tuning.

Hope this helps. I'm ok if it provokes any spirited rebuttals. Another cultural variation, like the credit thing: Mathematicians call each other idiots all the time (or so it appears from the outside) when all they did was mess up a sign. We're used to being corrected. I've noticed that people outside this tradition can be rather sensitive on forums to "competing evidence", even though our tradition here is primarily to set the record straight for all the lurkers who might actually believe us.

 

[HJaLTe]

First of all, I would like to say thank you for your replies so far. I know this issue is kinda hardcore (for me it is), so not that many know what to make of it.

Anyway - I'm a bit confused.

In Kris' article he posts this table:

Memory configuration table (ROG article by Kris)

... and calculates the latency in ns: TRD= tRDx1000/FSB (as Syzygies posted as well) - so far so good

... but how do you know what the tRD (NOT the TRD) is at any given FSB/DDR rating?

i.e. I run Corsair Dominator PC2-8500 at EPP 1066 MHz. My FSB is standard 333 MHz and the strap i on auto, but i asume it is 333 (hence the 1066 DDR rating). If we look at the table, my Default tRD is 10 and using the equation from above it gives at latency of 30 ns (I believe it can be a lot better than that).
If i OC my RAM modules to be somewhere between 1200 or 1300 MHz (1333 "given" by the 333 FSB) and increase my FSB to let's say 450 MHz, how will I find the correct tRD?
From what I can derive from the table, the tRD i not direclty linked til the FSB, but rather a combination of the FSB and the DDR rating.

I wouldn't say I'm mathmagical underchallenged, but I admit it's been a few years since I sat down with extensive calculus, so I might have "missed" the ansawer in your replies. However I still don't see where the tRD comes form, other than it seems like a given factor at a given DDR rating.


I DL'ed the memset program and it shows this:Memset
Is "my" tRD the highlighted one?? If it is, I can't see it in my BIOS. Other than that I can adjust pretty much anything in my BIOS.

//HJaLTe

 

[n7]

tRD is "Performance Level".

You're at 6, which looks about right for your FSB & RAM speed.

And i don't know how to calculate it...way above my head, especially at this time in the morning.

 

[HJaLTe]

I think I've seen the light - but I'm not sure!!!

Kris writes this: "Take note of the "Default tRD" column; you will see that each strap has an associated value (6 at the 200 strap, 8 for 266, etc."

So the tRD IS a given value fixed to the given strap you use.

i.e. My current settings are: FSB, 333MHz and the DDR rating 1066MHz = the strap is 333 = my current tRD is 10!!! (Well, if I can derive anything from Kris' table, this is the answer I have been looking for - Or is it???)

If I maintain the FSB at 333 MHz, but change the strap to 200, then my tRD should be 6 and my DDR rating will increase to 1111 MHz


[EDIT: After DL'ing the tRD calculater I'm bit confused again - in Kris' aforementioned table the Default tRD ranks between 6 and 12 (only even numbers), but in Tony' calculator it ranks from 5 to 8 (including uneven numbers like 5 and 7)

@ N7: Saying that my perfomance level is 6 complies with the tRD-calculater, Syzygies explains about - I see that. However, now I have 2 different tRDs: the ones from Kris table and the one from Tony's calculator - Unless each tRD in the calculator codes for the tRDs in the table)]

Maybe it's a trail 'n error process... *sigh*

//HJaLTe

 

[HJaLTe]

... Unless he uses "codes" for the "real" tRD's in the calulator.

Then I could imagine it would be something like this:

5=6
6=8
7=10
8=12

Or maybe that's just wishfull thinkin'

[EDIT: I just found something interesting. tRD = Performance level = Transaction booster. Depending on who you ask, apparently.
I have a transaction booster in my BIOS, so I think I might play a little with that (after using the calculator ofcause)]

 

[bryanW1995]

you can change the tRD with memset if your bios won't allow it. I don't know if that will also change your memory multiplier and other settings, however...be prepared to reset bios...

 

[flexy]

those are hyperinteresting articles.

Problem is only there is now ay for me to run ANY of those "recommended" settings,
eg. DDR2 1200 TRD/Perf lvl 5 ..you must kid me..i am glad if i can get to perf lvl 7
same with DDR 1000, CAS4 (!) abd TRD/Perf lvl 5 (!)

those people must have ultra-high end memory...i spend a few hours tweaking with OCZ XTC plats (basically reaperX) but there is no way to come even close to any of those settings as shown somewhere in that "best pick" table which is in the AT article.
This is on dfi X38/X48.

 

[Syzygies]

Originally posted by: flexy
those people must have ultra-high end memory...i spend a few hours tweaking with OCZ XTC plats (basically reaperX) but there is no way to come even close to any of those settings as shown somewhere in that "best pick" table which is in the AT article.

I have Reaper 2GB 800's, they're actually the same as the 1066's except for binning, and I got lucky, mine easily match the specs of the better sticks. Going to 8 GB, I ordered the better sticks; I'll retire the orginal sticks to a different build if they don't play nice together.

So yes, luck is involved in memory sticks like cpus. I'll eventually own four pairs and do my own binning, keeping the best pairs for the better build.

Yikes! I never realized until now that I was applying an X48 formula to a P35 chipset. It roughly works, though my P35 sometimes does better. I'm amazed no one pointed this out to me, I was about to point out that you were applying the article to the wrong chipset.

:roll:

I've heard (from leads starting in the OCZ forum) that the P35 actually handles DDR2 better than X48, which offers legacy support but prefers DDR3. The best DDR2 board for memory overclocking is reportedly the DFI LanParty P35-T2R, which also sports three PCIe-16 slots.

There's an "off by one" error in some formula behind all of the tables late in that article. I discovered this writing my own tRD spreadsheet, and using the article to check my work. My bad for not writing Kris. So you can cut yourself some slack, and back off a bit on the self-recriminations. Use the formulas, and your own spreadsheet, or at least work the numbers on scrap paper.