Atom Observations

I’ve been intrigued with low-power Intel Atom + NVIDIA Ion combinations for a while now and have recently put together a little net-top computer with a 1.8GHz Atom (dual-core + Hyperthreading) paired with an Ion2 plus 4GB RAM and the hard disk left over after I upgraded my PS3’s drive. The unit is a very small machine – bigger than a Mac Mini, but smaller than any minitower. It has Dual-Link DVI, so it can drive high-res displays, as well as a bunch of USB ports, eSATA, and a card slot.

Given the clock speed of this machine, I would expect it to perform pretty well. While it isn’t paired with a very new, fast hard drive, it seems to boot fairly quickly and is responsive (under Windows 7 64-bit). While I haven’t tried any CUDA tests on there, I ran Prime95 to test it and burn it in, and I’m surprised at how slow the Atom is. The 1.8GHz Atom takes between 8 and 10 times (not percent) longer than my 2.16GHz Core2Duo in my laptop for each of the Prime95 benchmark tests. My guess is that this is a very floating-point-intensive test, so it goes to show that the Atoms stink at floating-point computations (probably like the Cell in the PS3 stinks at double-precision computations). Perhaps FP is even emulated on the Atom.

So what? So the Atom is lousy at floating-point arithmetic? All people use them for is netbooks and set-top boxes, right?

Very true, but those netbooks and set-top boxes are sold as Windows (or Linux) machines that can run normal software, not special-purpose machines, like iPads and Android phones. Why does that matter?

Well, for years, FP speed has been getting better and better to the point that programmers were encouraged to use FP operations rather than faking it with fixed-point or making do with integer arithmetic. Games, graphics applications, and much more has become FP-intensive, since it is so fast on normal Intel and AMD processors, yet those programs will suffer if run on an Atom. Sure, we won’t be running equation solvers on Atoms, but this new reality is bucking a trend that has made programming easier (always a good thing) while providing good performance. While clock speed has never been a very good measure of performance, now more than ever, we need to be very clear that a 1.8GHz Atom is MUCH weaker than a 1.8GHz Core2Duo at some operations. Even older processors, like Pentium 4, will run rings around the Atom when doing floating point.

So I admire Intel’s ability to save energy with the Atom and to make it work well as a Windows host processor, but I am alarmed that they are willing to trade so much performance for power (though, frankly, Atoms are quite low power). Perhaps the power consumption is also 8-10 times (or more) lower than the Core2Duo in my laptop and is probably 50 times lower than that of a desktop PC, but the clock speed numbers are quite misleading when it comes to certain important kinds of performance. And that is bound to make some people unhappy.