[MASS] - MSI X58 Eclipse SLI
 

Thought it might be interesting to share some findings and thoughts of this motherboard. In fact, it's the same feedback I'm sending to MSI

As you might or might not know, yesterday I took part in the benelux finals of the MOA 2009 competition. Sadly, together with my team mate, I was only able to get second place, which means that I won't be competing in Munich. But, the overclocking session yesterday did allow me to learn the Eclipse SLI motherboard a bit better.

1) B2B timing
When I asked Janneke (of Mystar benelux branch office) if the latest bios was flashed, she answered yes ... but I didn't expect to see a bios with the B2B timing added :-). I quickly tested the effect of 0 and 12, but since it was a contest, I haven't had the time to look at it in detail. I have the board at home now and an 975 incoming, so I'll be testing it very soon.

In any case: timing set at 0 was perfectly stable.

Also very positive: we got the memory running at 1150MHz (2300CL9!) ... which is véry good :-). (if needed I'll check for the specifics of the kit used and send it to you)

2) Low-speed QPI vs. BCLK vs. PCIe
Since the processors used in the competition were 920's, the key of succes was getting the BCLK as high as possible. On air cooling, everyone was pretty much maxed out around 205-210MHz, which is similar to my findings when reviewing the board a few months ago. On LN2, and with more tweaking, two teams were able to bench at almost 230MHz BCLK.

For us, the key to get from 207 to 230MHz was to set the QPI frequency to low-speed mode and tuning the PCI-e frequency. The problem with this setting, however, is that it seems to have an effect on the PCI-e bandwidth as well. Now, I'm aware that lower frequencies will result in lower performance, but in this case, the performance was hurt so badly, I'm pretty confident it's not the QPI frequency only. When switching from full-speed (lowest multiplier) to low-speed the FPS in the 3D benchmark went from 1000+ to approx. 130, almost decimated.

In an (unreleased) article I wrote for Madshrimps, these are my findings for the influence of the QPI link speed on 3D performance:



In short: even with an 4870X2 (two cores on one PCI-e bus) the effect is incredibly small, almost insignificant.

My question is as follows: does the low-speed QPI frequency also have an influence on the PCI-e frequency or something else related to the PCI-e lane that results in a very low bandwidth? And if so: can this be changed or altered so that in low-speed mode, it's possible to do any 3D at full speed?

Does the low-speed QPI frequency also have an influence on the PCI-e frequency or something else related to the PCI-e lane that results in a very low bandwidth?

Yes,since the QPI is the link line between CPU and IOH, when change to low-speed, QPI only runs at 3.2 GT/s only. And it would also effect the PCI-E bandwidth as slowly and a lower benchmark result. Please see following pic for reference



can this be changed or altered so that in low-speed mode, it's possible to do any 3D at full speed?

Checking ...

There's still something I don't fully understand, though. As far as I understand, the standard frequencies and/or bus transfer speeds are:

• PCI-e 2.0 x16 => 16GB/s, 8GB/s per direction
• Slow-mode = 3.2GT/s => 12.8GB/s, 6.4GB/s per direction

So, the QPI link transfer speed is indeed lower than the PCI-e bus transfer speed and is the bottleneck for 3D performance. Now, in the overclocking session, we were running the following frequencies:

• BCLK = 207MHz
• PCI-e = approx 110MHz

So, the new transfer speeds would be as follows:

• PCI-e = (110/100)*16GB/s => 17.6GB/s
• QPI Link = (207/133)*12.8GB/s => 19.92GB/s

In this case, the QPI Link is no longer the bottleneck. Also, even if the QPI link frequency would be limiting the PCI-e bandwidth output, is the reasonably small bottleneck of 3.2GT/s capable of decimating the 3D scores?

Now, I have been browsing through the Intel datasheets and found this:

"The Intel QPI links will come out of reset in slow mode (66.66 MT/s) independent of the operational frequency. This is based purely on the reference clock (133 MHz) divided by four. Firmware will then program the Physical layer to the operational frequency,
followed by a soft reset of the Physical layer, at which point the new frequency takes over."
(~ http://www.intel.com/Assets/PDF/datasheet/320838.pdf - p28)

Apparently, the QPI-link can run in a 'slow-mode' at which the transfer speed is decreased to 66MT/s , which equals a 2.64GB/s data transfer speed. Is it possible that the "slow-mode" setting in the bios forces the QPI Link to work at this 66MT/s rate instead of switching to a higher frequency?

Assuming it indeed does, that would mean that the transfer speeds are:

• PCI-e: 17.6GB/s
• QPI Link = (207/133)*2.64GB/s => 4.11GB/s

Although this makes more sense to me, I'm not sure if it's indeed correct. The main problem here is that I don't have my i7 cpu to test all this at the moment. Tomorrow, a colleague will drop of my old 965, so this weekend I will test the influence of the QPI and everything related to it :-).

As promised I've spent a couple of hours testing the Eclipse SLI for the effect of the Slow-mode setting in the bios. I also did a few overclocking tests, but more on that topic later on this email.

Testing configuration:
- Intel i7 965
- MSI Eclipse SLI (2x)
- 3x2GB OCZ Gold series PC3 12800 - OCZ3G1600LV6GK
- GalaxyTech 9500GT OC

1) Slow-mode: performance results

In attachment you find the file "slow-mode tests.png", which is a screenshot of the test results. To measure the 3D performance, I used the 3DMark01 - Nature subtest.



The first two sets of data are to show the influence of the PCI-e frequency on the 3D performance. In both full-speed and slow-mode setting, the PCI-e frequency has a significant influence on the 3D performance: higher = better. We can also see the significant loss in performance when switching from full-speed to slow-mode: more than 50% loss in performance with a low-end 9500GT.

The third and fourth dataset are to show the influence of raising the QPI frequency through increase of BCLK frequency. As you can see: increasing the QPI frequency has nearly no effect on the 3D performance. The small increase can be explained by the increase of CPU/IMC/MEM configuration of which the frequencies are also affected by the increase in BCLK frequency.

The QPI frequency of 2.6GHz in the slow-mode data sets is the frequency reported by CPU-Z.

2) Slow-mode: effect on overclocking

First of all it's important to note I have tested two samples of the Eclipse SLI motherboard: one which I've been using since a couple of months after finishing my i7 round-up review, the second one is the one I used in the Benelux MOA competition. If needed, I can provide information regarding the serial number of the two boards.

On both motherboards, I experience the same effect of the slow-mode function: the maximum BCLK is higher than in full-speed mode, which is pretty obvious I reckon. The maximum BCLK on both motherboards is, funny enough, exactly the same:

230.42MHz - http://valid.canardpc.com/show_oc.php?id=589762

Note: I am using an C0/C1 processor, to be precise: an older ES sample of the 965. I haven't been able to test the effect of the new D0 revision, but as far as I can see, the results should be even higher. This 230.4MHz is one of the better results I've seen with C0/C1 samples so far, especially on air cooling.

The main issue is that this Slow-mode involves a significant decrease in performance, so we want to know the overclockability in full-speed mode as well.

3) Full-speed: overclocking results

First sample I tested was the one I recieved for reviewing purposes. Although the first series of tests proved 224MHz to be unstable in 3D situations, further testing got my upto 227.4MHz BCLK stable in 3DMark05; 228.4MHz was almost instant lock-up. The screenshot of the 227.4MHz stable BCLK is included in attachment as well. For CPU-Z, I could go upto 229.4MHz, 1 MHz less than in slow-mode.

Problematic is, however, that with the same settings, the second motherboard (the one we used in the overclocking competition) was only able to do 210MHz BCLK stable. In slow-mode, both motherboards performed the same, but in full-speed mode ... this one is much worse.

4) PCI-e frequency

It seems that beyond 105MHz, the increase in pci-e frequency doesn't help that much anymore for BCLK overclocking. On both boards I went upto 119Mhz, but the results were the same as at 105MHz.

A stock PCI-e frequency gives me exactly 222MHz BCLK on the best motherboard.

5) Performance beyond 222MHz BCLK

I haven't been looking into this all to much, but it seems that the clock-per-clock performance under 222MHz BCLK is higher than over 222MHz. If needed, I'll try to provide you with more data on the performance loss. Note that the loss in performance is not very big: maybe 10% maximum, in very specific benchmarks. My first guess would be a decrease in GPU bandwidth.

6) Questions

Not sure if these questions can be answered or not, be here I go:

- In the early days of i7, a lot of motherboards maxed out around 222MHz, but why that specific number?
- Is the PCI-e frequency derived from the BCLK frequency or another clock generator? I think the latter, but need confirmation
- What is the relationship between BCLK, QPI frequency and PCI-e frequency? If the PCI-e frequency is derived from a different clock generator, how can it affect the BCLK overclockability

7) Request
- CPU PLL voltage: would it be possible to add voltages options under stock voltage? So lower than 1.8v, because I found it to be quite helpfull in subzero overclocking sessions.
- Auto disable DRAM/PCI frequency: what does this option do?

8) Pictures


http://valid.canardpc.com/show_oc.php?id=589762

Increased max 3D stable BCLK on second Eclipse ... difference of 15MHz is a lot!