Hype(r)? . .

In a couple weeks, we’re going to hear a lot more from Intel and how it’s going to be so good for you.

At that time, we’re supposed to hear from Intel how much it will improve performance.

Let’s just say we’re pretty skeptical about hyperthreading doing a lot of good generally. No doubt it can and will help a few current problems a lot, but for more than that: Show me.

However, there is one common-used “application” that shows today how much good hyperthreading will do it. For that matter, you’ve been able to do this for the last few months.

What is it? SiSoft Sandra:

If you go to the CPU Performance Test, and futz around a little bit with the Reference CPUs, you can see this:


The latter represents a hyper-threaded PIV (that’s what SMT stands for). Quite a difference, isn’t there? Over a 20% improvement in ALU, and close to a 70% improvement in FPU.

Coincidentally (or not), hyperthreading will put PIV Sandra scores significantly above AMD’s Sandra scores. At 3.06GHz with HT, well, see for yourself. To match it, an XP would have to run at over 2.5GHz to match the ALU and forget about the SSE2 FPU.

Apparently, this boost will only be available for Intel chipsets.

A 20%+ boost at the same speed certainly would be a big, even huge deal, never mind 70%. But is this what we really have here?

For now, I think reading this article on Xeon hyperthreading to see real benchmarks on this. The benchmarks that liked hyperthreading got a 15-20%. Some weren’t affected at all. One program very near-and-dear to the outlaws out there these days actually got hurt by about 15%.

Seems to me that just about the most Hyperthreaded-optimized program out there is Sandra. This is not good, and greatly reduces the value of the Sandra CPU Performance Test as a general benchmark comparison between Intel and AMD.

If you think that an exaggeration, take a look at the articles written by Intel on the subject, and look at the benchmarks in the articles. Even Intel reports that most improvements were less than 20%, with none reaching near 70% (65% was the best for anything).

It’s one thing when a benchmark which uses a variety of programs uses them with the latest optimizations turned on. At least those programs exist, and they only constitute part of the score. All the programs that do not incorporate the technology won’t show improvement, and that will drag down the average level of improvement.

Here, though, this benchmark shows what looks to be every bit of improvement the program could get from a technology that won’t improve most current programs, and actually hurt some.

Yes, the Sandra people can come up with justifications for this, and frankly, they are caught between a rock and a hard place on this one. You can’t be half-pregnant on this issue; you either measure hyperthreading or you don’t.

What you need to do in the weeks ahead is pretty much ignore Sandra CPU Performance in any review, and look towards the other benchmarks to get a real idea as to what hyperthreading will do for (or against) you.


Be the first to comment

Leave a Reply