Hyper-Threading Performance Analysis
Published on 2002-09-30 13:57:42 By: Jim_

When I originally threw my Xeon box together and began running benchmarks for my Tyan Thunder i860 review, I didn't really give hyperthreading much thought. I proceeded to run our suite of both synethic and real-world benchmarks with hyperthreading enabled. Once the dust had settled and I compared the numbers with that of our 2000+ Athlon MP system, it seemed that the numbers were at the very least close in the majority of our benchmarks. The major exception was in our DivX encode of Antitrust, where the Prestonia's with HT enabled lagged behind our Athlon MP system by a whopping 10 frames per second. It was this rather shocking difference specifically, that motivated me to dig into hyperthreading a little more and see what potential performance impacts it was having; both positive and negative.

What is this "Hyperthreading"?

This isn't a new topic, so I'm not going to dig deeply into the technology behind Hyperthreading. If you'd like to dig into the technology behind hyperthreading, please refer to this technical document (care of Amazon International).

In brief, hyperthreading (initially referred to as Simultaneous Multi-threading or SMT) allows for a single physical processor to appear to the operating system as two logical processors. The operating system doesn't know the difference and feeds threads to each as if they were indeed separate physical processors.

In our case here at 2CPU.com, the power of two becomes the "power of four". Currently hyperthreading hasn't made it's way to the desktop Pentium IV processors (it's coming) and we haven't seen uniprocessor socket 603 boards to this point. I do know of at least one manufacturer who will be releasing a product like this shortly, and I will be testing it.

To the right, we see a screenshot of Windows Task Manager during a DivX encode. You see the four logical CPUs and they all seem to be at work.

To examine the performance impacts that hyperthreading has on our suite of benchmarks I enlisted Hooz to assist me in collecting the necessary data. I re-ran the benchmarks with 2GHz Xeons on my Tyan Thunder i860 and Hooz was kind enough to fire up his Iwill DP400 and run through the benchmarks again with his 2.4GHz Xeons.

Let's kick off the comparison with a couple of synthetic benchmarks you're accustomed to seeing run here at 2CPU.com, Sisoft Sandra and Cinebench 2000.

* All benchmarks were run under Windows XP Professional using Intel's i860 platform. (Yes, RDRAM for all those RAMBUS haters out there :-P)

There are no surprises in the CPU-related Sandra benchmarks. Sandra appears to be SMT-aware and the numbers back up that hypothesis.

The trend continues in the CPU Multimedia testing. Hyperthreading enabled is the order of the day with Sandra, it seems.

 
 [ Next to Page 2 ]