EVGA P55 Motherboard

The Winner of the March Blogger Giveaway is carpo93!

Congratulations to carpo93, and thank you to EVGA for sponsoring.  Stay tuned for April’s giveaway announcement later on today.

The idea of a low clock challenge is to provide a playing field for benchmarkers to prove their tweaking skills in a given benchmark. Setting specific limitations constructs the playing field and opens the doors to a lot of people to compete that might not have the newest or best hardware that a wide open competition would require. Overclockaholics.com recently had a low clock 3Dmark 01 challenge which locked CPU clocks to 4.2GHz and limited Nature Frames Per Second to 1200FPS for single card or 1400FPS for dual card entries. These two simple limitations really setup an array of possible winning combinations that require tweaking skills at both the software and hardware level.

Needing a bit of a break from pounding my head against the wall with a memory review, I took a couple days to get back in the 3DMark 01 tweaking seat – which hasn’t been sat in for some time – and tried my hand at the Overclockaholics.com 3DMark 01 Low Clock Challenge. Here is my story of how I came to winning the single card category last weekend.

There are two primary platforms that can be competitive in this type of challenge, using an Intel C2D on the 775 platform or the slightly less potent Intel i7 socket 1366 platform. The reason i7 is actually a worse a platform for 3DMark 01 is the fact that the Nature benchmark scores are terrible on the i7 platform at 4.2GHz compared to C2D. You will soon find out that I had plans to take care of that. Essentially though, I had both platforms setup ready to rock but my goal was to be competitive with the i7 platform using whatever means necessary. I began preliminary testing on the Intel i7/X58 platform with phase cooling on the CPU. The need for sub-zero cooling on the CPU will also be explained shortly.

With the GPU on air, it was apparent that being competitive just wasn’t going to be possible. At a minimum, 90K was going to be required to win this thing, at this point in the weekend, I think the top score posted for single card was around 89.5K already. Even with the CPU under sub-zero conditions allowing for a very nice uncore clock – which helps 01 scores tremendously – the Nature FPS were just too low to compete. That is when the big guns were brought out to help the little GTX260 play with the C2D boys.

A little rubber eraser, 1 KingpinCooling.com Tek9 4.0 Slim, a 50K ohm variable resistor, a couple shop towels, and this setup was ready to rock 01’s world. The variable resistor was the only mod done to the card which was done to bypass OCP. Other than that, this is just a straight up Gigabyte GTX260 216SP video card. Here is a complete list of the hardware used for the rest of the competition.

Over the course of two six hour sessions, this system was beaten, abused, and downright throttled. The GTX260 nicely surprised me willing to run through Nature at -140C. Normally these GPU’s will cold bug well before that without a special BIOS so either these cards had that special BIOS, or this card is just a freak of nature. Either way, the GPU clocks definitely helped the score and took the single card results for the competition to the next level. Here are a couple photos of the two bench sessions.

Most of the benching was done with the GPU at -120C and the CPU at -30C ~ -35C. The digital multi-meter is showing the resistance across the OCP mod which I set to 13.33K Ω. This allowed GPU voltage of 1.25v to be used at clocks in excess of 1000MHz. For the suite of benchmarks, GPU clocks were set to 1026MHz with shaders running at 2052MHz. The GPU memory clocks were most stable at 1215MHz but could creep close to 1300MHz. For Nature however, the GPU clocks were set as they are in the screen shot below, 1080/2160/1215. Here is the screen shot of my best – and winning – single card result.

Obviously this isn’t even the best this setup could pull off at 4.2GHz because Nature is far from maxed out. If I could get GPU clocks high enough to hit the 1200FPS limit in Nature, the overall score would have been at least 92K. Either way it didn’t matter as this was enough to win the competition and take home a very much needed KingpinCooling.com F1 EE CPU pot. I have been benching with a MMouse Rev3 CU pot for so long that the upgrade to one with more mass is guaranteed to help with multi-threaded benchmarks like Vantage and 06 with the i7 processors. To wrap things up, here are a couple photos of the setup during tear down. Plenty of snow was produced during the bench session, nothing like winter in July.

I would like to thank Overclockaholics.com for a great contest, the rest of the competitors for pushing me to go LN2 on the GPU, and KingpinCooling.com for the prize. All I can say is that this card is far from done.

A recent forum post out there in the middle of the Australian portion of the internet – AKA i4memory.com – led me to investigate the performance impact of a BIOS option on the EVGA X58 3X SLI Classified. The option labeled ‘MCH Strap’ would indicate that it changes the latencies of the MCH, like on previous Intel chipsets such as the P965, P35, X48 and so on. The only catch is that the Intel X58 chipset, does not have an MCH (Memory Controller Hub). The Intel X58 chipset is no longer responsible for the memory controller as Intel Core i7 processors now have a integrated memory controller – or IMC – on the CPU die.

This inherent difference in design of the chipset got me thinking, if there is no MCH, what is this setting actually changing? Apparently others were thinking the same thing as the question popped up in the i4memory.com forums. Needless to say, I took it upon myself to have a quick look, which you are reading right now…this very second, here at TechReaction.net.

Before we have a look at the testing, let’s go over the complete setup used for the testing, and discuss a little bit more about exactly what I am testing today, and why. First a couple photos of the setup followed by the list of hardware used.

This is primarily for Classified owners so I am just going to assume you are all familiar with the BIOS. On the main overclocking settings page is the option for the ‘MCH Strap’. There are a number of settings which are 2133, 1867, 1600, 1333, 1067, 800, AUTO, DRAM Ratio. On previous Intel chipsets, the MCH strap was responsible for various hidden chipset latency timings and a lower MCH strap would increase performance of the chipset, and subsequent memory subsystem. When adjusting the ‘MCH Strap’ setting on the Classified, the only visible changes are those to the memory sub-timings. A lower strap would result in tighter memory sub-timings. A higher strap setting and the sub-timings on the memory would loosen. Pretty straight forward…right?

One would assume so, but no one was sure in the i4memory.com forum discussion so I did this basic testing to see if there was anything being changed behind the scenes that we couldn’t see. The basic methodology is pretty straight forward. Set a high ‘MCH Strap’, then measure the systems performance and note the sub-timings. Set a low ‘MCH Strap’, then measure its performance the same way and also note the memory sub-timings that strap sets. Then simply set the higher ‘MCH Strap’ again, but this time manually adjust the memory sub-timings to that which the lower ‘MCH Strap’ set. If the performance is the same as the performance of the lower ‘MCH Strap’ setting, then we know there is nothing else going on that we can’t see.

I set the system up as outlined above. I decided on stock settings with the 2:12 memory ratio for no apparent reason really, I simply wanted to ensure I would have the ability to select a couple different ‘MCH Strap’ options. As it turns out, at these settings, the 1600 through to the 1067 ‘MCH Strap’ selection were the only ones available. That is why I only have results from those three straps. I also ran the AUTO option just to see what it would select. I used the Everest Ultimate and ScienceMark bandwidth test as well as a run of SPi 32M as that is the best way to see if any memory sub-system gains are had. I think that about covers the necessary details, so I’ll stop yapping and you can look at the results.

The results may appear almost identical but if we look close at the SPi times, there is a definite pattern there. From the higher ‘MCH Strap’ setting down to the lowest, we see the SPi time drop. This means we definitely have SPi gains from the lower ‘MCH Strap’ selections but that is primarily from the tighter memory sub-timings. Looking at the last result we can see that with the 1600 ‘MCH Strap’ selected and the sub-timings tightened manually to reflect the 1067 ‘MCH Strap’ selection, the result is the same time as the 1333 ‘MCH Strap’ selection. Considering there is only a 12 one-hundredths of a second difference between the 1067 and the 1333 ‘MCH Strap’ selection, I am going to conclude that the ‘MCH Strap’ option in the EVGA X58 3X SLI Classified BIOS is definitely just adjusting sub-timings and nothing more.

Obviously this wasn’t an exercise in exhaustive testing and I didn’t include any 3D testing but all indications of this setting were that it simply tightened and loosened the memory sub-timings. There is no reason to test 3D performance in my mind. Here is a full chart of all three runs of each setting, there was a reboot between each run.

To finish up I have included a sample from each ‘MCH Setting’ that was tested so you can see the memory sub-timings differences. If there are any further questions about this quick test, don’t hesitate to post a comment and I will do my best to answer. Thanks for reading.

The trip started innocently enough with a locally purchased i7 920 sporting the date code 3845B026. Soon after some quick air testing, however, things took a turn towards the crazy track with the idea to test a theory on whether a colder north bridge would assist with BCLK clocking. It turns out it didn’t, but that won’t prevent me from showing the vacation photos.

I had simply planned on multiple layers of Armaflex insulation tape for the NB pot but ran into a problem when I realized my roll of Armaflex tape was still making it’s way back from the GOOC 09 event in California this past weekend…so I had to improvise.

After some cutting of 1/2″ Armacell sheets, I had a nice little north bridge pot insulation sandwich. Paired with the mounted MMouse Rev 3 CU pot, I had my twin towers ready to rock. After looking at the north bridge insulation sandwich for a while, it gave me the craving for a clubhouse, so it was a couple hours and a few pints later that the benching began.

Complete Hardware Setup:
• EVGA X58 3X SLI Classified
• Stock PWM cooling / DetroitAC NB Pot / Swiftech MC14 / Enzotech SLF-1
• Intel i7-920 3845B026
• MM Rev3 CU Pot w/LN2
• Arctic Silver Ceramique
• Corsair Dominator-GT 1866 7-8-7 (TR3X6G1866C7GTF)
• ATI PCI Mach64
• Scythe Ultra Kaze 120MM 2000RPM 87.6CFM (DFS123812L-2000)
• Seagate 7200.9 80GB SATA II 8MB cache
• Custom N’lightnd Windows XP Pro SP3
• Corsair HX1000W

As mentioned, the cold north bridge didn’t assist with BCLK at all, as I got the same clocks as previously with the north bridge running at a steady 70C during a previous session. This wasn’t much of a surprise to be honest. At the 250BCLK point, the Classified simply needs fine voltage adjustments to eek the most out of a processor. It was still fun running the NB at -40C, and it seemed to handle those temps no problem at all.

Enough fooling around though, time to look at the results. First up is the highest BCLK I have reached with this combination so far. As mentioned, more is definitely possible with proper time put in fine tuning secondary voltages…but 250+ for validation is still plenty impressive

After fooling around with voltages for a few minutes to maximize the BCLK without too much effort, I turned my attentions to Super Pi, both 1M and 32M. I only ran 1M a couple of times and did get 250BCLK to run a couple of times but had freezes opening CPU-Z. This has happened to me before with this motherboard and requires the CPU pot to warm up a bit. I didn’t have patience because I wanted to move on to 32M and didn’t have much time to spend benching. Here are the two best results I have pulled with this combo thus far.

The first i7 920 under 7 minutes in 32M and if I am not mistaken, under 8 seconds in 1M. At the time of posting, the 32M is actually 6th globally ranked amongst all CPU’s on HWBot.org and this 920 absolutely crushes the 965 C0/C1’s I have tested. With a little more work on voltages and better memory, there are sure to be quicker times from this combo.

The EVGA X58 3X SLI Classified has been much heralded in the enthusiast and overclocking forums and for very good reason. This particular 965 was only able to clock to around 5040MHz on both the EVGA X58 3X SLI and DFI X58-T3eH8 for 32M SPi. On the Classified, however, I was able to push this processor almost 100MHz more at the same pot temperature of around -75C. This extra 100MHz mated with a solid set of Corsair Dominator-GT 3×2GB DDR3-1866 7-8-7 memory, got me below the 7 minute mark for the first time.

Complete Hardware Setup:
• EVGA X58 3X SLI Classified
• Stock NB, SB, PWM cooling
• Intel i7-965 Extreme Edition
• MM Rev3 CU Pot w/LN2
• Arctic Silver Ceramique
• Corsair Dominator-GT 1866 7-8-7 (TR3X6G1866C7GTF)
• ATI PCI Mach64
• Scythe Ultra Kaze 120MM 2000RPM 87.6CFM (DFS123812L-2000)
• Seagate 7200.9 80GB SATA II 8MB cache
• Custom N’lightnd Windows XP Pro SP3
• Corsair HX1000W

Using a fairly typical SPi setup, getting into the 6’s was a bit of a battle. At 5GHz it was evident that the setup had the chance to go sub 7 minutes but I had no idea how the CPU was going to clock on the Classified as this was my first session under LN2 working on 32M.

The only issue I had was getting a screen shot. After getting sub 7 a couple of times, the system insisted on freezing when trying to open CPU-Z for the screen shot. Even after an extensive copy wazza process and running 32M, the system didn’t like CPU-Z…which obviously wasn’t making sense. As it turns out, letting the pot temperature warm up to near -50C was what it took to get the screen shot. It only took about three Newcastle’s and a half dozen runs under 7 minutes for me to figure it out.

It isn’t the most efficient run out there and my copy wazza/32M OS needs some work but the goal was sub 7 minutes and that was accomplished. Thanks to EVGA for another fantastic motherboard and Corsair for a solid kit of memory. Next up is trying to get that time down by a couple seconds and break into the top 5 overall 32M on HWBot.org with this setup.