OC Contest Done Right: MSI MOA 2009 EU Finals

Tradeshow & OC events by massman @ 2009-07-22

Last weekend, Madshrimps took the plane to Munich to observe the European finals of the MSI Overclocking Arena 2009 competition. Instead of participating in the competition, our local OC Guru Massman crept behind the camera and reports from the EU Finals with scores, OC tips and lots of pictures!

  • prev
  • next

Problems and solutions

SuperPI 32M

As you already know from the previous page, the SuperPI benchmark is mainly stressing the processor and the memory subsystem. In terms of i7 terminology, this means that the following components have to be tuned for maximum performance :
  • Processor
  • Uncore (also known as memory controller)
  • Memory
As for the processor, the biggest issue with the Core i7 950 processor is that the turbo multiplier is 24, which is a so-called even multiplier. It's a well-known phenomenon that Core i7 processors do not overclock very well on even multipliers or to put it differently, they clock better with odd multipliers. So, everyone shifted down and worked with the 23x multiplier, but in order to get the processor up to speed a higher BCLK frequency was needed. As you can read on the Madshrimps forums, there seems to be quite a big variance between the overclockability of the motherboards, but fact is that for this competition most of the competitors were limited by the processor rather than the BCLK itself. For getting the highest BCLK frequency as possible, there are two tweaks that were used by the overclockers: PCI-e and Slow-Mode.

In the past, increasing the PCI-e frequency was only to increase the bandwidth throughput from the videocard to the Northbridge, but on Core i7, it seems that this frequency is linked to the overclockability of the BCLK. On the MSI, and in fact on many other boards, there's only great scaling up till 106-107MHz. Beyond this, there might be a little bit more stability, but not as much as going from 100MHz to 106MHz. Note that increasing the PCI-e frequency can lead to an unstable videocard configuration.

The QPI Slow-Mode is a different approach which doesn't always seem to help. Since this isn't really the place to be discussing this feature in detail, I'll keep it short: you're decreasing the QPI frequency by a larger factor in order to increase the BCLK overclockability. On some samples it helps a lot (+15MHz) on some samples not at all (+3MHz) ... that's really up to the board. Note that because you decrease the QPI Link spectacularly, the 3D performance drops significantly:

Madshrimps (c)

(QPI bus frequency is a possible bottleneck for the videocard since it connects the IOH (pci-e lanes) with the processor)

More on this topic you can find inside the Madshrimps forums: [MASS] - MSI X58 Eclipse SLI.

But, as already mentioned: the processor samples weren't that great, so most people only ran between 5.0 and 5.1GHz. As far the uncore, it's partially having luck, because as we also see on the processor core, there are certain multiplier issues. The problem is, however, that the multiplier issues aren't that easy to solve as on the processor itself, since there seems to be no general rule: certain processors (dis)like certain uncore multipliers, for other samples the bad and good multipliers may be different. So, if you are held back by the uncore frequency, it's sometimes worth to try a higher multiplier than a lower one ... but beware, although the uncore multiplier might be better, the increased frequency might require lower temperatures and/or a higher voltage for the uncore. And if that's not enough, there's also the problem that the coldbug and coldboot temperature of the uncore might be different from the one of the processor core. I think you can already understand why extreme overclocking is more difficult than what it looks like ...

Last but not least: the memory frequency. For this section, I could write a couple of pages, but I won't because I already did that in a previous article. One of the most important settings used in this competition was the 'Back-to-Back Cas Delay timing'. How it works, you can read here: Stabilizing your memory overclock on Core i7 platform - Back-to-Back Cas Delay Investigated. Most teams ran this timing at a value of 4 to 6, possibly 8. In the presentation before the competition started, overclockers were 'advised' to use a value of 31. Seems like MSI was trying to throw some people off track there!

3DMark03 / Aquamark

As for 3D, you need a well-clocking videocard and that cannot happen without the usual liquid nitrogen and voltage modifications. Overclocking legend Hipro5 spent about 16 hours trying to figure out how to modify the card for optimal performance and here are (some of) the modifications:

Madshrimps (c)

So, what are the modifications for:

  • Vgpu: increase the voltage for the core; possible issues: OVP
  • Total IMAX: remove the overcurrent protection on all channels
  • Channel IMAX: remove the overcurrent protection on each channel

    The two last may be a little bit confusing, because most of us are used to make just one modification for the OCP. However, after posing Hipro5 the question, and having received an answer, it's actually quite easy to understand why two modifications have been used.

    Madshrimps (c)

  • Imax-A = 30A
  • Imax-B = 25A
  • Imax-C = 15A
  • Imax-T = 50A

    In extreme overclocking situations, you will be increasing the current flow on channel A the most, since that's where you modify for increasing the voltage of the GPU. Now, since channel B and C will hardly increase, you can assume that they'll stay way below the maximum ratings, let's say 10A and 5A. So, that leaves up 35A for channel A if we only take the ImaxTotal into account. But, in order to use the maximum allowed current of 35A, you will have to increase the Imax-A since it's limited to 30A. So, in short, if you run into an OCP issue on this card, it's either because of too much current in total or too much current on one of the channels.

    (Note: values indicated are for example purposes only - they do not reflect the real values for this card)

    These few voltage modifications aren't enough, however, when you want to find the maximum limitation. We still have the OVP and Vmem modification. Before the competition, these didn't really seem that important, but afterwards it became clear that they were. As for the OVP, there's a hardware modification, but wasn't shared by Hipro5. The main issue is that on this card, the OVP kicks in at VID+125mV, which is quite early. The problem with the shared Vgpu modification is that it doesn't change the VID level, so you could only increase the voltage 125mV, which is not enough. So, in order to go higher, you need to change the VID, which can be done by either hardware modifications (not recommended), using the MSI software or use the software provided by Snot_aap. Each method will allow you to increase the VID and not to trigger OVP.

    As for the modification for the memory - you'll have to figure that one out for yourself as it has not been shared yet. It shouldn't be that difficult, though, since more than one team used the modification in the competition.

    The software tweaking part of this chapter can be held quite short: LOD + D3D overrider. For 3DMark03, many people used a LOD value of +15, whereas for Aquamark some used 0, some 2 and some 3. It's up to the reader to figure out the best setting for his hardware configuration.
    • prev
    • next

    No comments available.