AMD Phenom In-Depth Performance Scaling Analysis

Overclocking/ by massman @ 2008-12-30

With the upcoming release of the Phenom II we take a look at what we can expect by overclocking every aspect of our Phenom X4 9950 CPU. HT Link, Memory, Northbridge and RAW CPU speed scaling is compared in-depth to help you get the most out of your Phenom setup.

  • prev
  • next

HT Link frequency performance scaling

HT Link frequency:

Madshrimps (c)

Before we start analysing the obtained results, let's find out what the HT Link is designed for.

In Intel-terms, the HT Link frequency is quite similar to the FSB: taking care of the communication between processor and Northbridge, which features for instance the PCIe lanes, sata ports and audio support. The faster the link between processor and Northbridge, the more data can be transferred between both and, in theory, the higher the performance will be. Without going to much into detail, it's necessary to know that the current HyperTransport link is clocked at a frequency of 1800MHz, which is 1000MHz higher than the first edition of the HT link used on the s754/s939 Athlon64 platforms. Note that although the original HT Link frequency was set at 800MHz later revisions of the s939 motherboards had a stock HT Link frequency of 1GHz.

Test settings:
  • Core frequency: 2600MHz
  • HT frequency: 200MHz
  • Northbridge frequency: 2000MHz
  • Memory frequency: 400MHz (1:2 divider)
  • Memory timings: 4-4-4-10 2T Ganged

  • HT Link frequency (200 - 2000MHz*)

    (Click for bigger version)

    As expected, the HT Link frequency has little to no effect on the CPU and memory intensive benchmarks. Please don't jump to conclusions because you see a little bit of variance between the results: the differences are so small that they are in fact not significant. Do note the performance scaling effects in the Nature benchmark, which seems to benefit quite a lot from increasing the HT Link frequency. Well, actually we should say: seems to be affected quite heavily from decreasing the HT Link frequency!

    The reason why it's affected this heavily is actually quite obvious: the graphics card communicates with the processor through the Northbridge, using the HyperTransport bus. As we're using an HD4870X2 video card, which features two gpu's, the data output rate is quite high, meaning that there's a lot of information to be exchanged from processor to video card and vice-versa. The less high-end your video card is, the less of a bottleneck the HT Link will be. On-board integrated video cards, however, will be affected quite heavily, even though those are very low-end.

    Most of you probably wonder why I even tested this as the stock HT Link frequency is 1800MHz. On the one hand, because I'm a member of the third group of overclockers (see first page), on the other hand because this is one of the key features of Phenom II overclocking. AMD has been able to work around the coldbug issue, which has been quite an obstacle for our extreme overclockers for quite some time now, making it run absolutely stable at temperatures below -190°C, unseen on AMD processors using an IMC (integrated memory controller). The only issue is that the HT Link has to be set to 1.0 mode, which means using HT multiplier 5x. At first sight, this seems to be quite bad for the extreme overclockers, who run both cpu and gpu on LN2 cooling, but there are other factors to take into account.

    *: Because of the NB frequency of 2GHz, we were not able to test HT link performance scaling beyond 2GHz. The HT Link has to be equal or lower than the NB frequency at all times due to the design.

    • prev
    • next
    Comment from Kougar @ 2009/01/08
    Thought this was a great article...

    Just going to point out that while Phenom II increases performance, Phenom II decreased Uncore clocks from 2GHz to 1.8GHz. So AMD will be gaining that performance back with AM3 DDR3 processors launching soon, which will use a 2GHz Uncore.
    Comment from Massman @ 2009/01/08
    I'm looking quite forward to my PhII sample to compare the performance scaling.

    There's more to tell about the techy aspects of the P2 than one would think. Lostcircuits has some good insights: d=1