Some thoughts on Tref and DIY

[mad mikee]

There are a few threads w/ stuff about Tref and how it influences our OC. (AgentJCDenton in the Gskill BH-5 thread and the Tref thread come to mind). I have found the discussions useful.

However the frustration of not finding the 'ideal' TREF for many settings has led me to request this addition to the various biosi out there (Future DFI and ATI biosi especially )

1. First request
Replace the current tref selection w/ the following entries: 1.95 Auto, 3.9 Auto, 7.8 Auto, 15.6 Auto. These would calculate tref = FSB (integer) * the auto value and use the Truncated value (Accurate to 1 cycle) for TREF. Examples: Auto 7.8 x 275 = 2145, use that. Auto 3.9 x 311 (x 9 to get 2800 MHZ CPU) = 1212(.9 - toss that).

2. Second request
Allow US do the calculations and input our OWN value (saves bios programming, however, allows for GIGO tref values )

3. Third request - teh IDEAL
Do both 1 & 2 to so that we get a better tref for starters (1) and if the calc is slightly off for our particular piece of mem we can tweak it on our own (2)

BIOSI could have tables for the trefs for above refresh rates from 150-350 FSB and let us handle anything outside of that on our own

[Cornelious0_0]

I'm completely in favor of this.....the tref selections seem to be the only thing in the BIOS for my Ultra-D that seem oddly out of order and in need of help.

[Jimbo Mahoney]

From my understanding, if the Tref is too high, the memory isn't getting refreshed enough and risks losing data. If the Tref is too low, it's getting refreshed too often, causing instability due to power / heat issues.

I've done an excel spreadhseet where the user inputs their memory stick size and their memory speed and it calculates a theoretical ideal Tref.

It's here:

http://homepage.ntlworld.com/james.waynewright/Tref.xls

Let me know if it needs changing, or feel free to make it into a webpage as I don't have the 1337 5k1ll5 required to do that...

Does it work for people?

[mad mikee]

Now if only we could put in those ideal values

Would you please share methodology that led you in making this tool (being the inquiring minds we all are )
Tanks mIkeE

[Jimbo Mahoney]

OK,

It started here:

http://www.dfi-street.com/forum/show...98&postcount=3

Now I also know that 128Mbit memory modules (256MB) like 15.6us refresh and larger modules (512MB) prefer 7.8us.

That's where the first user-specified variable comes into the spreadsheet.

The second one is just based on the fact 1000/clock speed = absolute time period of 1 cycle.

So, for 200 Mhz = 5ns

Thus, as you ramp up the clock speed, you need to increase the Tref to keep the refresh period at 7.8us or 15.6us.

Get me?

[Jimbo Mahoney]

OK, I've updated the spreadsheet so users can input their stick's default refresh period if they know it.

[hovo73]

Thanks for the spreadsheet, but I am still unclear as to which values are "high" and which ones are "low". The values seem to be in complete disorder and I really don't know what "ramping" up really means?

For example, I have 2x512 sticks and I am trying to get 252 MHz, according to the spreadsheet, the ideal TREF value is 1968.75. How do I get that value, or the one closest to it?

EDIT: I don't know the default refresh period, how can I get it? (OCZ EL Gold BH-5)

[Jimbo Mahoney]

OK, basically, the Tref in the BIOS is the number of cycles before a refresh occurs.

Since the higher the speed of the memory (in Mhz), the less time each cycle occupies.

Thus, in order to maintain the refresh period of the memory (normally 7.8us or 15.6us), you need to adjust the Tref as you adjust the memory speed.

Increasing the clock speed = decreasing the cycle length = need to increase the Tref setting in the BIOS.

This is all theory at the moment, hence the need for testing.

As far as the spreadsheet goes, just choose the value(s) closest to the theoretical value and test.

In principle, Everest can give you your RAM's refresh period, although it is telling me my G.Skill 4400LE is rated @ 2.5-4-4-8 @ 200 Mhz, so who knows if it's accurate or not

[WeStSiDePLaYa]

think anyone can write a little proggy to do this as not everyone has excel, also if the proggy only selects values that are OPTIONS in the bios. thanks! anybody else think its high time we get one?

[pershoot]

Thanks for the spreadsheet, but I am still unclear as to which values are "high" and which ones are "low". The values seem to be in complete disorder and I really don't know what "ramping" up really means?

For example, I have 2x512 sticks and I am trying to get 252 MHz, according to the spreadsheet, the ideal TREF value is 1968.75. How do I get that value, or the one closest to it?

EDIT: I don't know the default refresh period, how can I get it? (OCZ EL Gold BH-5)

refresh should be 7.8us for those modules.

[Cornelious0_0]

The spreadsheet is working fine for me.....nice little thing to have on hand....and once I take the time I want to do some proper CPU burn-in with my Venice I'll see how much it helps fine-tuning my Tref as opposed to just setting it to 3072 irregardless and that's it.

[Jimbo Mahoney]

OK, well it's working for me.

I also e-mailed it to a friend who reported positively:

"OK just tried it

Before: tref 1168 (166/7.8 in A64 Tweaker),

Everest latency 36.5, Read 7913, Write 3345

According to your spreadsheet, 512MB at 300 requires Tref of 2344. So tried 2336 (nearest value), which is 166/1.95 in A64 Tweaker (but 7.8 us in DOS Memtest A64 Options!).

Gives latency 36.5, Read 7928, Write 3369.

So a slight improvement."

[mcnbns]

I messed aroung with Tref last night a bit, and finally got 260 MHz 2-2-2-5 stable. I used Tref 2048 through the BIOS, and A64Tweaker reports it as 133 MHz 3.9us, if that's of any use to anyone.

[mrlobber]

I messed aroung with Tref last night a bit, and finally got 260 MHz 2-2-2-5 stable. I used Tref 2048 through the BIOS, and A64Tweaker reports it as 133 MHz 3.9us, if that's of any use to anyone.

According to Jimbo's calculation, the theoretical ideal value should be 2031, and obviously, 2048 is the closest to it. Did you find that value by yourself? If yes, that would be another confirmation this calculation is working

[Jimbo Mahoney]

Oops. Edited

[Cornelious0_0]

So then you ARE supposed to use the next highest value that's higher then the ideal value.....or lower then it?

[g0dM@n]

think anyone can write a little proggy to do this as not everyone has excel, also if the proggy only selects values that are OPTIONS in the bios. thanks! anybody else think its high time we get one?

I can probably do that if I can remember stuff back from 5 years ago with Visual Basic.

I can try to write a program if someone can tell me how to make it *.exe.

[THunDA]

Thanks for the spread sheet thingy.. Ill try this later..

[Jimbo Mahoney]

So then you ARE supposed to use the next highest value that's higher then the ideal value.....or lower then it?

Try both. I would use the one which is closest, whether higher or lower.

Too high and too low are both bad for different reasons.

[g0dM@n]

Nice OC, Jimbo... is your Venice holding you from running higher mem?

Also, does anyone know the exact uS (microsecond, or cycle) value of the new TwinMOS Speed Premium (1A4T --> BH5)? I'm at work, but I'm curious if it's 7.8uS or not.

[burningrave101]

Thanks for the spread sheet thingy.. Ill try this later..

Ya know, once you figure out your refresh rate all you have to do is multiply that times the speed your RAM is running at and that will give you the tREF your shooting for. You get the same value if you calculate using the spreadsheet. I know cause i tried it :p.

[flexy]

hey this is a VERY interesting thread.

Problem: I dont have excel and the excel-viewer doesnt allow changing values.

I am running two 512MB sticks and everest says default refresh is 7.8uS.

But now i am running at 291 FSB !

Could you explain (in easy terms)..how to calculate the TREF out from
the HTT/FSB and modules (7.8uS) i have ?

This was NOT that clear in that thread.

[amrgb]

I have the exact same FSB as you.

Using the formula you should use a Tref of 7.8*291=2270. Following the theory you should try the Tref's near this value. Try 2336, 2560, 2592, 2064, 2048.

In practice, I have better results with lower Tref's, like 648, 780 and 908.

For me there is really one way to find the optimum Tref. Try them all with Superpi 32M as jot down the number of loops you can do before failure. For me it's much better than memtest. As an advantage you can easily change the Tref with A64 tweaker. The problem is to test the tref's that cannot be chosen in the tweaker.

[Jimbo Mahoney]

Simply:

1 / memory speed

then divide your refresh rate by that number

e.g.

1 / 200 Mhz = 0.005

7.8/0.005 = 1560

Although amrgb's method is way easier! silly me

[flexy]

Simply:

1 / memory speed

then divide your refresh rate by that number

e.g.

1 / 200 Mhz = 0.005

7.8/0.005 = 1560

Although amrgb's method is way easier! silly me

yeah i figured that out...


but something is weird...do we have to take DDR2 into account ?

I remember the default tref for my sticks (200 HTT) is 3170 or something for 200mhz.....and this would make sense if they use 400 (!) (DDR2) and not the 200 HTT...otherwise they would NOT come up with a default of 3170 but, as you said...something in the 1500 range !

Short: Do we have to multiply the result x 2 ??????

(look at the default tref values in bios if you set everything to auto and default HTT)