Difference in RAM overclockability based on CPU freq?

[jimbob91]

First off, system specs:

Epox EP-8KDA3J with Sempron 3100, 512 meg ram (2x256 Mushkin level 1 2-3-2)

The Sempy will hit 2.38Ghz at stock voltage stable, and 2.47 Ghz at 1.5-6 volt.

What I noticed:

I could reach 230 FSB stable with the Muskin RAM @ 2.7v, using prime and memtest. Much beyond that, no go. On this board, I have heard, the 2.8v setting doesn't work, so that doesn't help overclocking efforts. Also, loosening timings didn't seem to help me reach any higher.

Next, set RAM to 166, and got to 255 FSB (or so), before getting unstable, but the RAM is only running at 208. I can push towards 260 FSB, which is getting the RAM around 212, but that's it. Why isn't 230 FSB possible like original? Is this the mem controller onthe chip not handling the increased CPU freq?

At 265 FSB (2385Mhz), I have to run 133 mem, which puts me well under 200Mhz on the RAM. I tried using A64tweaker to set the RAM to 150, which should put the RAM at 198 Mhz, still under it's spec, but that causes instant or near instant lockup.

Any ideas why I would get less and less out of the RAM as the CPU freq increases?

Great forum ... have learned a lot in here!!

 

[VirtualLarry]

Makes perfect sense to me - higher CPU core clockspeeds require more power, which leaves less available for driving the integrated DRAM controller signals stably. In fact, I tend to wonder if perhaps that is part of the problem with some of the Winnies not running Prime95 stable at stock - due to the process-shrink and the lower vcore, I wonder if perhaps the DRAM controller isn't getting enough power under heavy loads, and they might have to re-design it slightly, just like they re-did the TBred-A and turned it into the TBred-B with an additional metal layer, to allow it to clock higher stably.

Note that simply increasing the vDIMM won't help this issue - that supplies more (core) power to the DRAM chips, allowing their internal logic to potentially clock faster, but the memory-control signals are driven by the CPU's DRAM controller.

As an aside, I kind of wonder if the CPU-integrated DRAM controller has programmable signal drive-strength values, much like many system chipsets do or did. That used to be the thing to tweak to get overclocks to work successfully, on AGP systems with non-async PCI/AGP/FSB speeds - adjusting the "AGP drive strength value" on many chipset would afford greater stability on the AGP signals at higher speeds. This setting was independent of the GPU/VRAM's voltage or frequency.

 

[BoleCailey]

By the ocz voltage booster. Thats what I need to get.

 

[Duvie]

Yep....I have seen that as well...

My geil stick would run 230fsb 1:1 at cas 2 no less for 460ddr yet once I got up to 330 would not run 270 with a 166 divider for 460ish ddr at cas 2.5....At 330 is had issues running a 133 divider with cas 2.5.....This was irregardless of vdimm.....

I did notice opne thing the higher divider required more vcore...meaning stable at 133 divider needed more vcore (+.03) to get 166 divider to work on same speed again regardless of vdimm....It has to be cause the memory controller in ondie...

 

[jimbob91]

Might be an interesting test ... I'll try giving the Sempy a little juice earlier that it needs,and see if I can run the memory higher.

Checked out the OCZ booster, but I don't think it is compatible with this board.

Thanks for the ideas!

 

[jimbob91]

Well, I did some more testing.

At 260FSB x 9 =2340Mhz Prime95 stable for 24 hours. Memory is running at 218-19 via 166 mem divider (1/11) as the BOIS displays, but by calculation at CPUZ say about 213.

I did quite a bit of testing at 263FSB. This only raises the memory a little (220 at Bios, 215 by calculation) At that speed, tried raising vcore, no difference. Lowered it back down and loosened timings to 2.5-3-3-10, not stable. Loosened to 3-4-4-10, not stable. Tried 2T command rate, not stable.

Then, set back to 2-3-3-6 timings that are standard for these modules @200. I tried upping the chipset voltage, and initially, thought I had something. Instead of immediate P95 failures, ie within minutes, it would last an hour or so. So it made some difference, but not enough.

Keep in mind, these Muskin modules ran at 230+ when the CPU was around 2.1 Ghz. Anyway, I'm just running it at 260x9 for now. I'm not sure another 50 -125 Mhz is worth the decreased bandwidth.