Taking BH-5 to its limits - Project Log

[felinusz]

The goal of this project, is to take my BH-5 based system memory as far as it will go with complete stability. In this thread, you will see no unstable results that are not labelled as such - the distinction between complete stability, peak screenshot speeds, and peak useable (bencheable) speeds shall be made very clear.


To accomplish this goal, I will be doing a variety of modifications, tweaks, and cooling enhancements to the memory itself, and to my motherboard.

I am doing this mostly for the fun of it, as well as out of curiosity. Hopefully, by the end of this project, some of my questions will have been answered through firsthand results and through firsthand gains...

Is our BH-5 memory held back by the IC itself? Is it held back be heat as a result of extremely high overvoltages, or is it held back by the older JEDEC-Standardized memory PCBs that BH-5 commonly comes on? At how high a VDIMM will the memory continue to see stable overclock gains? Can I crack a 300 MHz 1.5-2-2-3 CPU-Z screenshot with this machine, and with this memory?



The memory on the workbench is my Corsair XMS PC3500 BH-5 based memory. This is a 2x256 meg set of RAM that I use for absolutely everything - gaming, benchmarking, work, internet browsing. I have been running this RAM at 3.78V of VDIMM for say-to-day use since I got this computer onto it's feet.

To start out with pre-project results, without any modifications the memory is completely stable at 273 MHz 1.5-2-2-3, with 3.78V of VDIMM, and a single 80mm fan blowing over the sticks.

Our testbed is my A64 machine, a 3700+ San Diego and a DFI nForce4 motherboard. The memory is actively cooled by a single 80mm fan which is installed onto the memory slots with a homemade fan clip.




I ordered a set of 16 system Memory IC sized copper heatsinks about a week ago, they arrived on Friday.

Before the RAMSinks were installed on this memory, the sticks peaked at 273 MHz, using 1.5-2-2-3 timings, for 24/7 stability across memtest86, Prime95, and 3DMark. The memory peaked at a benchmark-useable (unstable) speed of 275 MHz. A VDIMM voltage of 3.78V (Measured with a multimeter) was required for these overclocks.

After the RAMSinks were installed, the memory saw 5 MHz in gains to it's 24/7 stable overclock - 278 MHz 1.5-2-2-3 is now completely stable across memtest86, Prime95, and 3DMark with the same VDIMM voltage of 3.78V.

This gain, while marginal, is impressive in the sense that it took the memory past it's previous peak-useable speed, but with complete stability. In the grand scope of things, 5 MHz is not a whole lot, making the RAMsinks quite an un-cost-effective means to get more from your memory - but signifigant gains were present.

The peak benchmark useable speed was increased to the same degree that the peak stable speed was increased - the sticks are now benchmark useable to 280 MHz, and can be run through EVEREST and Sandra as far as 282 MHz.


An other effect of the RAMsinks, is that the memory now appears to continue seeing gains when run at higher VDIMM overvoltages. Previously, a VDIMM above 3.78V (with 3.9V being the next step my motherboard made available to me) would adversely affect the stability of any given memory overclock - the RAM would lose an edge when given more voltage.

Apparantly this was due to the increase in heat as a result of overvoltage outweighing the benefit of the overvoltage itself - with the RAMSinks installed the memory now sees gains when run at 3.9V of VDIMM (My DFI nForce4 motherboard overvolts the VDIMM that is set within the BIOS - 3.7V in the BIOS becomes 3.78V when checked with a Multimeter, 3.8V in the BIOS becomes 3.9V when checked with a Multimeter.).

While I have not extensively tested the Windows-based stability gains at 3.9V (I am wary of running the RAM at such a high overvoltage for extended time periods, although it is the memory PCB rather than the BH-5 ICs that I am concerned about), the memory generates fewer errors per pass in memtest86 when run at an unstable clockspeed. At 3.9V, 279 MHz is stable in memtest86 test #5 for half an hour, which is as long as I ran it for. A 1 MHz gain from a .12V overvolt is very, very, very bad voodoo . Generally, when diminishing returns have kicked in to such a strong degree it is a good indication of an unsafely high overvolt.

Here's a shot of the heatsinks installed on the sticks (The heatsinks were actually slightly smaller than the memory ICs... ThermalTake has once again produced an imperfect product ):

 

[felinusz]

To me, the really interesting thing about these results is that they confirm my firm belief that if properly cooled, the BH-5 memory IC should be able to handle voltages in excess of 3.7V for extended time periods.

Later this week, I am going to do a VDIMM voltmod to my motherboard, in order to get around the board's overvoltage of BIOS VDIMM; a voltage modification will allow me to set VDIMM between the currently available values of 3.78V, and 3.9V. The lack of VDIMM 'fine-tuning' as a result of the motherboard's overvoltage prevents VDIMM overvoltage anywhere between 3.78V and 3.9V - a huge .12V span that needs to be filled in order to get a better idea of performance gains.

After fixing the board for better VDIMM control, I will also be attempting capacitor modifications to these memory sticks, which should theroretically net another 3-6 MHz on the memory's stable overclock. With the extremely high signal voltage and enormous current draw that the memory PCB is being subjected to (I assume that this memory PCB was designed for a maximum of ~2.9V of VDIMM at a relatively conservative current draw), the capacitors that are already on there quickly become inadequate.


And here are some "glamour" (THESE SPEEDS ARE NOT STABLE, OR EVEN USEABLE FOR BENCHMARKING) write/latency/read results run at 282 MHz. Please keep in mind that the memory's peak for stabilty is a good 4 MHz below these numbers.

In addition to these preliminary numbers, I can run pretty much any benchmark on request that anyone would like to see . I will also try to aquire a digital camera, in order to take a photograph of 278 MHz 1.5-2-2-3 passing memtest86 stability testing.... it is a glorious sight for sore eyes .

 

[jesse]

Wow thats interesting! I didn't know that having heatsinks would acctually improve ram stability and ocability like that! Nice

 

[Zebbo]

Nice to see a little gain there, but why the bandwidth is so low?? Especially write bandwidth.

 

[RedDragonXXX]

Hmmm, I'd like to put some of these on my 6800U Ramsinks.

 

[makshim]

Hmmm, I'd like to put some of these on my 6800U Ramsinks.

OCZ BGA Copper RAMsinks are your friend.

 

[felinusz]

Nice to see a little gain there, but why the bandwidth is so low?? Especially write bandwidth.

That is a very good question. I'm running extremely tight timings, 1T (CPC ON), and in DC. Bandwidth is however proportionally low, in the sense that at a lower memory speed the bandwidth is lower than what you see above .

Keep in mind that I have not tweaked my timings for numbers, rather for stability. The timings which I am using are the same which I use day-to-day, to squeeze as high a frequency as possible out of the memory sticks, without compromosing performance too much. The Latency test in particular can be pushed a step or two lower.

 

[Otter]

Interesting. Do you have a way to measure the temperature of the IC's?

 

[felinusz]

Yes I do .

I checked temperatures at the base of one of the RAMSinks. Idling, at 279 MHz and with 3.83V of VDIMM, the heatsink gives me a 24C readout. 10 minutes with a memtest86 Test #5 load at 279 MHz and 3.83V of VDIMM, the heatsink gives me a 33C readout. The sticks have a ~30cfm 80mm fan blowing directly overtop of them.

My Ambient is 20C, as measured with the same temperature probe device that I used to check the RAMSink temperatures.




In other news, I got the VDIMM and VTT voltage modifications done on my motherboard yesterday, and have done some brief work fine-tuning the VDIMM voltage. At 3.83V of VDIMM, the sticks are memtest86 test #5 and test #8 stable at 279 MHz for 5 hours each. 280 MHz will run cleanly throuh test #8, and will run a good ~10 passes of test #5 cleanly before it starts to error. 280 MHz is very, very close to being completely stable.

285 MHz 1.5-2-2-3 with complete stability, under 3.9V, is my current goal for this project.


All DFI nForce4 motherboards have a rather serious VTT fluctuation issue under certain load conditions, such as memtest86 test #8. The Drive Strength values can offset this somewhat, but the fluctuation is still present. VTT has a very tight tolerance, a VTT droop below reference (1/2 of VDIMM) can adversely affect stability. I am currently running my VTT slightly hot in order to compensate for the fluctuation (The VTT fluctuates between 1.92V and 1.95V with my current setup - VDIMM is set to 3.83V) - a slightly hot VTT does not seem to hurt stability.

My SMD Capacitors arrived today, I will be VDIMM Cap modifying these memory sticks tonight if I have the time (about ~130 SMD caps and ~260 solder joints are involved in this... all of the VDIMM regulation caps on the memory PCB need to be "doubled up"). I expect to see a good 3-6 MHz in stable gains from the cap mods.

I also intend to fabricate a new RAMFanClip with 2 80mm fans to replace the single 80mm fan one that I am currently using. Twice the airflow and air pressure over the sticks should net another 1-2 MHz or so.


Here are pictures of the VDIMM and VTT voltage modifications that I did to my motherboard:

 

[Sucka]

Interesting read, and very well written.

How much does raw CPU speed affect Everest? I wouldn't mind seeing some Sandra and Super PI.

How about a A64Tweaker shot as well, or are all the settings in your sig current with this batch of results?

I'll be watching for the progress.

 

[felinusz]

The settings in my sig are what I am using now, although I think they'll be needing a big change, because my Sandra efficiency is around ~78% - something is seriously wrong with my current setup. I've set up my timings for stability, I'm running the sticks in the Orange slots, DC, 1T, tight timings listed in my sig .

My Sandra numbers at 279 MHz are compareable to some UTT results that I have seen, at 260 MHz.

I ran SPi 1M at 279x9, the picture and checksum is below . Even this result is quite poor for the clock frequencies - it is on par with an A64 at 260x10, 2-2-2. Granted, I have done no OS tweaks to improve the time (~32S).

279 MHz appears to be SPi unstable, 2M would not complete without an error. I'm going to do some more work on tweaking my timings, and see if I can:

~ Figure out which setting is killing bandwidth
~ Get bandwidth up to par for this speed
~ Get 279 MHz SPi stable

 

[Gautam]

Awesome work as always Ian, you da man! I've only dreamt of getting mine up that high. And before you ask, yes I've tried up to around 4.1v.

Something is up with that PI though, should be 28sec at worst, I'd imagine. Perhaps the memory's instability is manifesting itself in the time and hurting it.

 

[felinusz]

Thanks G!

Something is up with that PI though, should be 28sec at worst, I'd imagine. Perhaps the memory's instability is manifesting itself in the time and hurting it.

That could well be the case, but the low Sandra efficiency tells me that my timings are off somewhere - likely my tREF or my Bypass values . It's a crying shame, because it likely means that my SPi stable numbers will be lower than my memtest numbers.

I need to go through them again, tweak each one and see how they affect efficiency in Sandra . I've been doing most of my testing with memtest, Prime, and the 3DMark Lobby test.



I have bought a cheapo digital camera off of Greenman100 - I'll have pictures of 278/279 MHz 1.5-2-2-3 passing memetest86 T5 and T8 in a few days (Hey, I gotta back up the stable results, right? )

 

[Sucka]

Yeah, those scores are certainly low for those speeds/timings. I realize that CPU speed plays a role, but your scores are as you said on par with UTT at ~260ish. I wonder how much the 512mb plays in that

Here's some Sandra with my UTT @ 264 although i don't know what that extra MHz and 512mb has on the score. Just something to compare to i guess. Regardless amazing overclock

 

[speed bump]

I would say that the 2x256mb is holding you back on some of the benches A64's love dual channel one gig. Also becuase of the inagrated ram controller A64s really need a nice shot of cold to max out the FSB. Also i'm not sure about the BH-5 record but somebody at XS did 306 screenies bith BH-6.

 

[Janus67]

that is very well written and is some good learning material for myself. nice post felinusz, good luck maxing out!

 

[Otter]

Thanks for the temp data, Felinusz.

I am currently running my VTT slightly hot in order to compensate for the fluctuation (The VTT fluctuates between 1.92V and 1.95V with my current setup - VDIMM is set to 3.83V) - a slightly hot VTT does not seem to hurt stability.

I'm not sure what you're saying here. Do you mean you increased Vdimm just for the sake of stabilizing Vtt? Or... you let the mosfet's responsible for Vtt get warm up on purpose?

Interesting pics. How do those mobo mods help?

 

[Zebbo]

You may want to try out Tras: 6, for me with the Redline stuff, 6 and 7 gives more bandwidth than 5 does. But that cannot explain that bandiwidth which is freaking too low. Are you sure you have enabled Bank Interleave?

 

[NinjaZX6R]

This is extremely impressive to see such a large amount of work put into this felinusz. It makes me want to put in my old Kingston BH-5 again just to test. I was definately surprised to see a 5mhz gain with simple ram sinks, as well as better response to voltage. What I noticed on my bh-5 is that anything past 3.8v does not yeild any better results. Like you mentioned, the over voltage was dimishing any overclock gained from it. VERY good point that I think some people fail to realize. The sky is NOT the limit by any means as far as memory voltage is concerned. None-the-less, thanks for taking the time to educate a lot of people on how to get the most out of their BH-5. I think between the few of us who still run BH-5 over UTT, we may create another crazy demand for BH-5!!! Well, maybe....

-Collin-

 

[felinusz]

Otter

I'm not sure what you're saying here. Do you mean you increased Vdimm just for the sake of stabilizing Vtt? Or... you let the mosfet's responsible for Vtt get warm up on purpose?

Interesting pics. How do those mobo mods help?

I have installed a modification that keeps tracking VTT 'properly', but allows me to swing it up or down by ~0.01V.

ALL DFI nForce4 motherboards have a VTT fluctuation issue in certain conditions - VTT will fluctuate by about ~0.03V. It doesn't sound like much, but since VTT has a very tight tolerance, it becomes a rather serious issue. VTT is supposed to be exactly 1/2 of VDIMM - a 0.03V fluctuctuation is a lot.

By "running VTT hot" I mean that I am overvolting VTT slightly, in order to compensate for the fluctuation. By running VTT above reference, when it fluctuates and droops, it does not drop below where it should be. A VTT that is lower than it should be will hurt stability, a VTT that is higher than it should be ("hot") does not adversely affect stability . The fix is a "brute strength" one, the VTT fluctuation is still present. Some folks over on XS are working on a true fix that will completely remove the VTT fluctuation.

The VTT mod has not increased stability that I can notice. It has however fixed my board's VTT from going below reference, which cannot be a bad thing.

The VDIMM mod has not improved stability, at the same VDIMM voltage. Yet it is still extremely useful, because it allows me to do fine-tune adjustments to VDIMM - my board overvolted the BIOS VDIMM before, resulting in very, very rough VDIMM adjustment invrements.

When playing with such high VDIMM overvoltages, for 24/7 use, fine-tune adjustments are pretty important, and have been very useful so far

Gautam

Something is up with that PI though, should be 28sec at worst, I'd imagine. Perhaps the memory's instability is manifesting itself in the time and hurting it.

Are you sure? I looked around for compareable results, and I found a 1M result from OPB at ~2900 MHz, with his BH-5 running at 293 2-2-2 he managed 27S in 1M - http://www.xtremesystems.org/forums/showthread.php?t=52245

NinjaZX6R

This is extremely impressive to see such a large amount of work put into this felinusz. It makes me want to put in my old Kingston BH-5 again just to test. I was definately surprised to see a 5mhz gain with simple ram sinks, as well as better response to voltage. What I noticed on my bh-5 is that anything past 3.8v does not yeild any better results. Like you mentioned, the over voltage was dimishing any overclock gained from it. VERY good point that I think some people fail to realize. The sky is NOT the limit by any means as far as memory voltage is concerned. None-the-less, thanks for taking the time to educate a lot of people on how to get the most out of their BH-5. I think between the few of us who still run BH-5 over UTT, we may create another crazy demand for BH-5!!! Well, maybe....

-Collin-

Thank you! The RAMSink results were fascinating to me as well, I think that this IC is partially just held back by heat. Bigtoe told me to "treat the BH-5 IC like you would a processor" (paraphrased) - it appears that he is quite correct.

Zebbo

You may want to try out Tras: 6, for me with the Redline stuff, 6 and 7 gives more bandwidth than 5 does. But that cannot explain that bandiwidth which is freaking too low. Are you sure you have enabled Bank Interleave?

Right now I am using a reserved tRAS value of 3, which actually seems to net my memory some small gains (as measured by the number of errors per pass per half hour at an unstable clockspeed in memtest86+ test #5). DRAM Bank Interleaving is Enabled. Thanks for the advice dude, I'll try a tRAS of 7/6.

I did find out what was killing bandwidth, and lowering efficiency - it was my multiplier. Apparantly a 9X multiplier is terrible for efficiency. At 10X, the numbers are closer to where they should be, as is the efficiency . My Processor is being pushed to it's limits at 10x279 - it is highly unstable at 2790 MHz. I also squeezed a 2M run, although 4M is still unstable at 279 MHz.


Below are some pictures of badnwidth at 279, with a 10X multiplier. Please note that 279 MHz is not completely stable - it is SPi unstable, but passes memtest86.

Here are the new numbers (The "eve" pics are the new Everest numbers at 279 MHz):