The hypothesis

We have been mailing back and forth with Gigabyte, because in our initial testing 200MHz BCLK with IGP enabled did not work either. In fact, we were, as well as Anandtech, stuck at around 165MHz BCLK using the configuration as stated on the previous page. Although the initial results were kind of a set back, we continued our research and with the help of Gigabyte finally managed to get the system going at higher BCLK frequencies. However, instead of being satisfied that we can make a screenshot of a stable configuration ... we want to find the exact cause of the problem and an appropriate solution for it. So what's the problem?

As you can imagine, the problem is located at the level of the IGP frequency. Although it can be overclocked fairly well on air cooling - 1.1GHz is not impossible - it does impose a limitation when increasing the BCLK frequency to roughly 165MHz. And that's perfectly understandable, as long as you keep in mind that the IGP frequency is, as well as any other frequency related the CPU, relative to the BCLK frequency. To solve this IGP frequency issue, many mainboard manufacturers have already implemented a bios option that allows users to change the IGP clock frequency. On the mainboard we've tested, this bios issue can be found in the overclocking section.



Now, against all expectations of both the end-user and, apparently, the manufacturer, it seems that this bios option does NOT set the graphics core frequency to a fixed value. What it does is very simple: it changes the IGP base frequency. As none of the current applications allow me to correctly measure the graphics core frequency, I have to test this theory using a less elegant methodology, but ... at least one that works.

The two key words of this methodology are "performance scaling". During my test sessions, I've used the 3DMark03 benchmark to see how much performance was delivered by the IGP frequency. Since the performance level of this 3D technology is quite low, the influence of CPU and QPI frequency are neglegible, if kept within small margin, when comparing results. Here is the list of results:

FORMAT: "BCLK/IGPclk= score"

200/500 = 5166

200/666 = 6543

200/700 = 6989

150/auto = 6567

165/auto = 7116

Given that I'm working with a Core i5 661, auto in this case equals "900MHz". The scaling is not entirely right if we only consider the frequency set in bios, so we have to come up with a different formula to calculate the IGP frequency. After a bit of research, we came up with this one:



Using this formula, we can recalculate the resulting IGP frequency

FORMAT: "BCLK/IGPclk ~ Real IGPclk = score"

200/500 ~ 750MHz = 5166

200/666 ~ 1000MHz = 6543

150/auto ~ 1015MHz = 6567

200/700 ~ 1050MHz = 6989

165/auto ~ 1115MHz = 7116

As you can see, the performance scaling result match the IGP clock frequency or, to make it a bit more scientific, the data fits the hypothesis. For argument's sake, let's try it the other way around as well: I make a prediction based on the hypothesis, run the benchmark and see if the prediction is correct. I ran these the following BCLK/IGPclk combinations:

154/626 = 4892

160/600 = 5110

166/578 = 4958

180/532 = 5061

182/527 = 4851

190/505 = 5040

200/480 = 4817

210/456 = 4717

220/435 = 4943

(Clickable - Screenshots are a selection of the results and cannot be analyzed separately)

All the same resulting IGP clk of ~ 721MHz assuming the above formula is correct, all pretty much the same frequency. I will add the screenshots of the results to the forum thread of this article. Note that a 900MHz IGP will score roughly 6000 to 6200 points ... the variance as seen in above list is not significant enough to disqualify the hypothesis.