Findings and some thoughts

First of all, it's not that easy to present a well-based conclusion (using valid arguments) about the overclockability of the Gulftown retail B1 revision after having spent only a 15L of LN2 in roughly 4 hours time. In fact, you must understand the consequences of what I am doing here: I'm trying to generalize the overclockability of one particular setup and form a conclusion that suits the entire range of Gulftown-based platforms, so I already ask you in advance to forgive me if the findings and conclusions turn out to be incomplete or inaccurate.

Moving on. As you might (or might not know), being into extreme overclocking requires me to be part of several extreme overclocking forums and websites to share information about overclocking. Many of my exploits are posted on forums a bit before they are turned into articles, because I prefer to hear the feedback from other users to check if my findings are somewhat accurate. When writing an article, however, a small forum post isn't enough, so I'll try to make it a bit more elaborative.

Keep the hero A0 or is it done with B1

The biggest question I have to tackle is one that have kept many extreme overclockers busy for the past couple of months: is the A0 revision really so much better than the retail B1? Comparing the results from the previous page to those you can find on HWBOT, you'd come to the conclusion that this is indeed true. However, we need to make a few sidenotes to this comparison before we jump to conclusions. First of all, as with the 32nm-based dual core codenamed "Clarkdale", the only variable that really seems to help the overclockability is the temperature. Much more than increasing the voltage, it seems that decreasing the temperature is beneficial for both getting the system more stable and reaching higher clock frequencies. What do I mean by this?

Increasing the Vcore helps a tiny bit to improve stability in the upper region of the "max clock zone", by which I mean that in case of almost 5.8GHz stable, adding a bit more Voltage will help to stabilize this 5.8GHz. However, adding a lot of voltage will not yield you 200MHz, certainly not at -110°C.

So, as already said, the operating temperature is the key for reaching high clock frequencies. And that's where the A0 versus B1 revision becomes a bit problematic. Because the A0 was an engineering sample only designed for system builders or mainboard manufacturers to check if their PCB design can handle a Gulftown, it does not have any security features built-in. For instance the lack of thermal or overcurrent protection comes in quite handy when overclocking and allows you to push the processor way beyond what it's designed for. Now, moan all you want, but adding these security features is a vital process of designing a new architecture, especially when they hit retail. Without these protections, normal end-users have an increased risk of damaging their €1000 processor even in normal test environments ... I don't think that's what we want. The upside of this story, for extreme overclockers at least, is that some A0 samples have no coldbug whatsoever, meaning they'll be able to operate at -196°C in theory, -185°C in practice. Given our "the temperature is vital"-finding, that's a huge advantage.

Overclocking B1 - Some tricky situations

Here's the tricky part however. With the B1 sample, I was able to reach 5.6GHz Wprime 1024M stable at only -110°C ... which is not that much worse than A0 at the same temperature and voltage level, according to several A0-overclockers. So, and this is where I might be wrong, I do believe that B1 revisions WILL match the early A0 overclocks, but only if you find one that is capable of running the very low temperatures. Not many have been tested so far, so it's quite difficult to come to a conclusion, but we certainly hope this will be happening in the future. In any case, I most certainly invite Intel to meet up with overclockers and check if there's any correlation between stepping and lowest bench-able temperature. Preferably myself *wink*.

In addition, I have to add that A0 revisions tend to die quite quickly and have an inferior memory controller on board.

How did the UD7 do?

So, was this session all happy-happy pink roses? No, because I experienced several other problems, not all related to the processor. For one, it seemed that the Gigabyte X58A-UD7 was not capable of handling either the extreme current or the extreme temperatures. Let me explain.

While I was indeed reaching 5.6GHz 12-thread stable, sometimes for no apparent reason the motherboard shut down and locked itself in the "88" position, meaning "unable to start POST". Powering off the PSU, let the board rest for a couple of seconds and powering everything back on certainly solved the problem (even bios settings were saved), but the sudden shutdown combined with the inability to commence post is a clear indication of current problems.

As for understanding the extreme temperatures issue, it's important to know the difference between coldbug and coldbootbug. The coldbug is the absolute temperature at which the component fails to operate. The coldbootbug is the absolute temperature at which the component fails to boot up. The coldbootbug is generally a bit higher than the coldbug, because it seems that components have no problem continue to work at a given temperature, but have more problems going from idle to load at a given temperature. To give an example: the coldbootbug may be around -100°C, but when in windows the coldbug may be -125°C. The issue in this case is that the coldbug and coldbootbug behaved in a very atypical manner. It's a bit hard to explain, but in general it works like this:

- Coldbug => board shuts down completely and doesn't commence POST (88)
- Near coldbootbug => board starts POST, but fails regularly (C1-C2-C3-shutdown-C1-...)

In this case however, the board shut down at exactly -123°C, but it didn't behave like a coldbug should as it returned to C1 directly, not 88. Dropping the temperature even more, I found the coldbug-like behavior to appear at -155°C, which makes me to believe that the processor is actually capable of much lower temperatures and, thus, higher frequencies. This is something we need to figure out in the next couple of weeks when testing other mainboards!

Last couple of things ...

Another subject I have to tackle is an issue with VTT voltage scaling, but this depends from sample to sample. I was not able to set a VTT voltage higher than 1.5V or the mainboard would refuse to boot. Although the memory controller quality of the B1 revision is a lot better than the one of A0, with 1.495V it's unlikely we'll see clock frequencies in the 5GHz range. Other people report the same issue, but at different VTT voltage levels.

One last thing for the extreme overclockers under us: this architecture consumes liquid nitrogen like crazy, especially when stressing all 12 threads in multi-threaded applications like Wprime 1024M. At first sight maybe a bit strange given that the TDP of a 32nm 6-core (980X) is equal to that of a 45nm 4-core (975), namely 130W, but you have to know that in contrary to the Bloomfield platform, this architecture is actually scaling with voltage. Whereas our beloved quad cores maxed out around 1.6 to 1.65V, the Gulftown can easily go up to 1.8 to 1.9V, which is exactly where the difference is located. The issue of the temperature was noticeable even while using the solid copper Kingpincooling Dragon F1EE, known for its mass. To give you an example: when I start the Wprime benchmark at -110°C and a container 1/2 filled, the temperature was raising 0.3°C per second. Even with a fully filled container, the temperature drop was still noticeable as ~ 0.15°C increase per second.