[linuxkernelnewbies] EclipseOC | OC Guides | HyperTransport Link Investigation

[Peter Teoh]

http://eclipseoc.com/index.php?id=6,47,0,0,1,0 Introduction Due to some questions I have answered recently, I concluded it would be a good idea to do a formal look into the effects of an overclocked Hypertransport Link on a system. Before I get into the results, I will first explain what this so called "HT Link" does, and more importantly, what it does not do. The entire premise of the HyperTransport Link is to provide a high bandwidth & low latency data pathway for the CPU to communicate with the rest of the system. Currently, the only CPUs in the wild with this HT link (instead of the old aging EV6 bus) are the K8 based AMD CPUs. If you wish to learn about the nitty gritty, the HyperTransport Consortium has a wealth of information at their website. Now, onto what it really does without the fancy terms: Connects the CPU to the Chipset. Allows CPUs to talk to eachother in MULTI-SOCKET configurations. What it does not do: Connect the CPU to the RAM. Connect two cores in a dual core CPU. This task is left for the System Request Interface that all K8's have. It's important to keep in mind what this implies. For most things, the bandwidth provided by the HT Link is overkill. For a s939 K8, the HT link is specified to operate at 1000mhz, double data rate for 2000mhz effective, 16bit width each direction for a 32bit effective. This all adds up to a staggering 8gb/s of bandwidth. Consider the PCI-E 16x link, which has a maximum bandwidth of 4gb/s. This is the only thing that will come anywhere close to saturating the HT link right now, but one must keep in mind that video cards do not normally use up all of the available bandwidth. So, all this points to the HT link spec being a bit overkill, no? Let's see if it is or not. First off, let's have a graph of all the data collected. I ran the system as follows: Opteron 146 @ 2.6ghz 2x1gb TeamGroup Cronus @ 260mhz - 3-3-3 timings Geforce4MX 440 @ 300mhz core, 317mhz memory 1x80gb WD, 4mb cache 2x250gb WD SATA, 8mb cache (not in RAID) I use the 250gb drives for storage, I ran HDtach on one of them to determine if the HT link speed has any noticable effect on hard drive performance. 3dMark 2001SE was used to get a rough idea of video card performance. I would have used a newer, more powerful bench that stresses the HT link more, but my lowly GF4MX was not up to the task *tear* CPUMark and SuperPI 1M were run to get approximations on the amount of CPU processing power available. SiSoft Sandra was used in a similar way to see if memory performance changes with HT link speed. Chart with all results. As we can see, the numbers do not change a whole lot. The only one with any noticable downward trend is the 3dmark score, which shows that even even my lowly video card likes having bandwidth and low latency. I wish there was a benchmark to directly measure the bandwidth and latency of the HT Link to compare it to these semi-real world tests. I feel it would be also be a handy tool for system diagnostics, as there have been some reports of performance reducing if the HT Link speed is pushed too far. I presume this is due to the packet nature of the link, the bad packets are thrown out and sent again, reducing the effective performance of the link. Conclusion The HT link appears to effect one thing in it's current state: Video subsytem performance, and even there it's not a tremendous change until we get down to the very lower limits of the speed range that is available. It's been said before and I'll say it once more to drive the point home. The HyperTransport Link speed has a very small effect on system performance. I invite you to discuss this article in our forums. Registration is quick and painless ;)