Overclocking 939 CPU's with ram at high fsb's

[Tony]

Guys

Im seeing a lot of posts where members are jumping on the ram because they can't hit the fsb's they want, issue is it is probably the memory controller. Now i have been testing Winchesters, Newcastles and Clawhammers for a few days now and the differences in ram clocks are huge between them.

Winchesters will boot and run benches at some high speeds, but they are far from prime stable. I have now decided to only show benches here of my dimms after an extended period in prime as CPU clock speeds have a huge effect on how the ram clocks. I can run 3d mark at 2.65GHZ but prime needs 2.35GHZ or so to be stable, thats just one example.

Newcastles run a LOT hoter than Winchester's, they usually have a pretty good memory controller but you may find you have to run more voltage to the cpu and the ram to clock high. Memory timings will be different to Winchester also, TREF and TRFC have a huge impact.

Clawhammer seems to have the slowest memory conroller, it also requires the highest overvolt on the dimms to run high fsb's. Again TREF has a huge impact as well as TRFC. Timings in general will be close to Newcastle but again different to Winchester.

So as you see, if you have a Newcastle you may not be able to try someones tweaks on the ram if they were using a winchester etc. you have to be prepared to run different timings which may mean the ram is slower dependant on what CPU you have. A good memory controller will make good ram shine, a poor one will make the ram look a dog.

I hope you all understand what im trying to get across to you all here.

[bachus_anonym]

Hi bigtoe,

Good move, bud.
I've always been wondering why people don't include e.g. at least some large PI calculations not mentionig about Prime95 runs, when testing memory sticks. Also, moving to Windows memtest rather than DOS based could also be a good idea. As i've seen cases when RAM is stable in DOS Memtest but totally unstable in Windows enviroment.
Showing only Sandra or Everest screenies or SuperPi 1M just doesn't cut it anymore. And furthermore, doesn't prove stability of the RAM at certain, high speeds.

Looking forward to some new benches from you, bigtoe.

best regards,

[Tony]

The issue is i see many running high speeds etc but most use phase change or high end water to gain the stability they need on the cpu to get the ram stable, most readers here are on high end air cooling and i have a feeling their expectations on stability are a little higher than they actually are in reality.

As i said here my Winni does 2.65GHZ air cooled and will run all the 3d i shoot at it, it bombs in prime in under a min....far from stable as you see.

The purpose of me posting this thread is to try and show people they have to remember A64 clocking is heaviliy dictated my the memory controller on the cpu, without a good one your ram is doomed.

[blinky]

thanks for the heads up

[Alex08]

what about those week 40 and before winchesters that failed prime at stock speeds. is that due to the mem controller? i mean i can run superpi 32m stable at 2.7ghz 270 2.5-3-3 with tccd's

[Tony]

what about those week 40 and before winchesters that failed prime at stock speeds. is that due to the mem controller? i mean i can run superpi 32m stable at 2.7ghz 270 2.5-3-3 with tccd's
I have a week 31 here thats fine with prime as long as i don't get over excited with its overclock, i know some wouldn't run prime at all so you would have to move onto other apps that may show up the instability.

32M super pi is pretty good as is memtest for windows.

[uranium]

How do you rate the memory controller on the FX 55 Tony?
Its week 46 2004 if that makes any odds
Thanks :)

[jikdoc]

i've been doing lots of ram testing on ocz vx, 4200el, 3200rev2, gskill LE, twinmos utt. the differences i see between the sticks seem to lie more in the range of normal variance than in the manufacturers. on to the topic, with the dfi board, i've recently come to appreciate how much tref, trc, trfc and max async timings affect stability and bandwidth. the old guide on memory timings over at dfi-street needs to have an addendum for nf4.

question on tref. i noticed on the new dfi bios, default tref was changed to 15.6. is this optimized for ocz vx, tccd or ...?

i noticed on another post you commented on settings for vx, i was hoping you could give some tips on a64 timings for tccd at >300 htt. i'm having a great deal of trouble getting anything higher than 310. (i'm on fx55 vapo)

[NapalmV5]

LOL, so gf4 was right afterall.

[Tony]

i've been doing lots of ram testing on ocz vx, 4200el, 3200rev2, gskill LE, twinmos utt. the differences i see between the sticks seem to lie more in the range of normal variance than in the manufacturers. on to the topic, with the dfi board, i've recently come to appreciate how much tref, trc, trfc and max async timings affect stability and bandwidth. the old guide on memory timings over at dfi-street needs to have an addendum for nf4.

question on tref. i noticed on the new dfi bios, default tref was changed to 15.6. is this optimized for ocz vx, tccd or ...?

i noticed on another post you commented on settings for vx, i was hoping you could give some tips on a64 timings for tccd at >300 htt. i'm having a great deal of trouble getting anything higher than 310. (i'm on fx55 vapo)

If your running 2x512 of TCCD at 310 your already doing a lot better than you should be, The load on the memory controller at this speed is huge, many only post benches using 256's as the load is a lot lower. I would say your probably maxing the CPU there, a new E die FX will probably help you go higher.

TREF at 1.95us 166 was a huge issue on the VX, I found that setting 15.6 or 7.8 usually cured the memtest errors but lost bandwidth over some of the more agressive settings. i have actually found a tweak for VX that gains 100MB's in memtest but you need damn fine modules and a good memory controller to run it.

[Gitsum]

You mention the E die FX. Tell us more.

[Sentential]

Ive been thinking about this very thing myself. Any ideas as to why the memcontroller on most Winchesters is FUBAR? Could it be because of a direct dieshrink? Or is this why the wattage has increaseed on the new E0s because they modified the memorycontroller in some way..

all very interesting stuff.

[slayer899]

TREF at 1.95us 166 was a huge issue on the VX, I found that setting 15.6 or 7.8 usually cured the memtest errors but lost bandwidth over some of the more agressive settings. i have actually found a tweak for VX that gains 100MB's in memtest but you need damn fine modules and a good memory controller to run it.

excuse my ignorance on this topic but which of the two will give more bandwith and is the preferred setting for the TREF? mine is now at 200 3.95 us, but if 166 1.95us will give me better performance then i can always try that....btw i have TCCD ram, OCZ rev.2 :)

[Tony]

You mention the E die FX. Tell us more.
Its coming

[Tony]

Ive been thinking about this very thing myself. Any ideas as to why the memcontroller on most Winchesters is FUBAR? Could it be because of a direct dieshrink? Or is this why the wattage has increaseed on the new E0s because they modified the memorycontroller in some way..

all very interesting stuff.

The issue is the cpu's appear to clock real well, the actual core can do the speed, the memory controller can not. I have actually thought of testing with the ram clocked down to see if this cures the issue but im back on a newcastle so it will have to wait.

I think we will have to make a decission with regard to winchesters unless you get a stellar one with an an awesome memory controller.... MHZ or ram speed.

[Tony]

excuse my ignorance on this topic but which of the two will give more bandwith and is the preferred setting for the TREF? mine is now at 200 3.95 us, but if 166 1.95us will give me better performance then i can always try that....btw i have TCCD ram, OCZ rev.2 :)
Test it and see, its quick and easy with A64 tweaker.

[cantankerous]

what is this tweak you found for 100mb more on memtest? Stellar sticks and mem controller or not I know I'm curious to know what this tweak is. Care to share?

[Highland3r]

Having MASSIVE OC issues with this 3500 and gig of PQI turbo atm...
memtests fine at 260, hell, its clean at 270 (2.5,3,3,7, 1t) just cant get the cpu to run at those speeds at a decent clock... Backed down to 255 to *try* and get something out of it, but still no ball....
SuperPi 1mb, 3dmark, everest tests all pass fine at 280... prime wont, and anymore than a few 3dmark loops = no go :(
Seems there *is* an issue somewhere, cpu's run 300 5:4 @ 2640, w/ 1*256 fine, 2640 270 5:4 2*256 fine as well... Just wont run this with a gig :(

[Tony]

what is this tweak you found for 100mb more on memtest? Stellar sticks and mem controller or not I know I'm curious to know what this tweak is. Care to share?

Play with TREF and idle cycle limit at 32 or 64

[cantankerous]

ok thanks!

[eva2000]

Play with TREF and idle cycle limit at 32 or 64
Tony how does TREF in bios translate to values in A64 Tweaker i.e. 3120, 4708 = what in AMD64 Tweaker 200mhz 3.9 ? any formula to work it out without trying each setting and constantly rebooting into windows ?

[Lidocaine]

Interesting thread.

I've been trying to tighten my timings as tight as possible and remain 100% stable, and with the plethora of timing options I was able to get 6072Mb/s (http://216.77.188.54/coDataImages/p/Groups/251/251420/folders/190829/1432242SreenshotofDFIbandwidthefficiencyatdefault. JPG) on SiSandra at default clocks and timings of:

DRAM Freq. Set = 200 1:1
CPC = Enabled
CAS Latency(CL) =2
RAS to CAS(Trcd) = 2
Min RAS# Active time(Tras) = 5
Row Precharge Time(Trp) = 2
Row Cycle Time(Trc) = 7
Row Refresh Cycle Time(Trfc) = 10
Row to Row Delay(Trrd) = 2
Write Recovery Time(Twr) = 2
Write to Read Delay(Twtr) = 1
Read to Write Delay(Trwt) = 2
Refresh Period (Tref) = 4708
Write CAS Latency(Twcl) = Auto
DRAM Bank Interleave = Increase
DQS Skew Value = 0
DRAM Drive Strength = Auto
DRAM Data Drive Strength = Level 4
Max Async Latency = 7
Read Preamble Time = 5
Idle Cycle Limit = 16clks
Dynamic Counter = Enable
R/W Queue Bypass = 16x
Bypass Max = 7x
32 byte Granularity = Disable(8burst)

ADV. CHIPSET FEATURES:
CPU Spread Spectrum = Enable
PCI-e Spread Spectrum = DownSpread

These DFI boards are extremely efficient and hold voltages better and utilize memtimings unlike any other NF3 or NF4 board I've owned....My other boards cannot generate 607xMb/s on default clocks.....

I could definitely see how just one wrong timing could throw a stick in someones spokes ESPECIALLY when overclocking.

[Skip]

my gskill LE at stock clocks can run cas 1.5 you should try and see if it works.

[Zebo]

try and show people they have to remember A64 clocking is heaviliy dictated my the memory controller on the cpu, without a good one your ram is doomed.

Are mem controllers different in same CPU iteration? eg winchs can have different mem controller? How can this be? Arnt they all made the same way? :confused:

[Lidocaine]

my gskill LE at stock clocks can run cas 1.5 you should try and see if it works.

Thanx for the suggestion.........will give it a go. ;)

[OCFreeze]

question on tref. i noticed on the new dfi bios, default tref was changed to 15.6. is this optimized for ocz vx, tccd or ...?

i noticed on another post you commented on settings for vx, i was hoping you could give some tips on a64 timings for tccd at >300 htt. i'm having a great deal of trouble getting anything higher than 310. (i'm on fx55 vapo)
That's supreme with two sticks X512Mb .With EPOX 8KDA3+ and A64 3000+/2G/ and 2X256Mb A-Data Vitesta DDR566 with TCCD on Vdimm=2.8 and 11-4-5-3 i'm having a great deal of trouble getting anything higher than 295 , to be honest. :rolleyes:

But with only one stick everythng works synchronous at 2T CPC with 11-4-5-3 I made 332.7 ,also I was with HT-->X4 and Vdimm=2.8V


http://forum.abit-usa.com/attachment.php?postid=497931

[Zebo]

Are mem controllers different in same CPU iteration? eg winchs can have different mem controller? How can this be? Arnt they all made the same way? :confused:
BUMP

[Zedex]

Tony how does TREF in bios translate to values in A64 Tweaker i.e. 3120, 4708 = what in AMD64 Tweaker 200mhz 3.9 ? any formula to work it out without trying each setting and constantly rebooting into windows ?

take the 200 and multiply it by that 3.9 and you'll have your Tref value in the BIOS. 200*3.9 = 780 ;)

Edit: Seems it doesn't seem to be that way for all Tref values, just happened to work out for a few I've seen available in the BIOS and posted in screenshots of A64 tweaker on this forum.

[Chr0n1c]

Thx bigtoe.

[Zebo]

Guys

Im seeing a lot of posts where members are jumping on the ram because they can't hit the fsb's they want, issue is it is probably the memory controller. Now i have been testing Winchesters, Newcastles and Clawhammers for a few days now and the differences in ram clocks are huge between them.

Winchesters will boot and run benches at some high speeds, but they are far from prime stable. I have now decided to only show benches here of my dimms after an extended period in prime as CPU clock speeds have a huge effect on how the ram clocks. I can run 3d mark at 2.65GHZ but prime needs 2.35GHZ or so to be stable, thats just one example.

Newcastles run a LOT hoter than Winchester's, they usually have a pretty good memory controller but you may find you have to run more voltage to the cpu and the ram to clock high. Memory timings will be different to Winchester also, TREF and TRFC have a huge impact.

Clawhammer seems to have the slowest memory conroller, it also requires the highest overvolt on the dimms to run high fsb's. Again TREF has a huge impact as well as TRFC. Timings in general will be close to Newcastle but again different to Winchester.

So as you see, if you have a Newcastle you may not be able to try someones tweaks on the ram if they were using a winchester etc. you have to be prepared to run different timings which may mean the ram is slower dependant on what CPU you have. A good memory controller will make good ram shine, a poor one will make the ram look a dog.

I hope you all understand what im trying to get across to you all here.

Do you think it's CPU speed or HTT speed causing premature winny failure? If it's CPU speed it does'nt matter what version you buy 3000/3200/3500 winny's should all crap out at around similar CPU MHZ overclock. If it's High HTT causing failure..Might then we opt for 3200's and 3500's since they allow lower HTT because of thier higher multi?

[Larethus]

That's supreme with two sticks X512Mb .With EPOX 8KDA3+ and A64 3000+/2G/ and 2X256Mb A-Data Vitesta DDR566 with TCCD on Vdimm=2.8 and 11-4-5-3 i'm having a great deal of trouble getting anything higher than 295 , to be honest. :rolleyes:

But with only one stick everythng works synchronous at 2T CPC with 11-4-5-3 I made 332.7 ,also I was with HT-->X4 and Vdimm=2.8V


Yes! I am having a problem along these lines as well. I have a winnie 3200+ wich clocks pretty nice (2.7ghz), NEO2 platinum, OCZ PC4200 EL TCCD/brainpower. With both sticks in there it is not memtest stable at ddr533 (stock). Yet with only one stick it is stable at ddr533 and beyond. Damn flaky memory controller :mad: Running it @ 10x270 in sync with both sticks will produce instability. I can run it at 9x300 with the 166 mem divider and it seems stable enough (ddr500)
I don't really want to buy a new cpu till these E revision chips come out, then it will be an FX-57. Guess those of us with a crappy memory controller have to rough it out for now....