Thursday, April 9, 2009

Comparing PC performance

A couple of weeks ago, I was trying to get a handle on the performance of an algorithm I had implemented a couple of decades ago. I wanted to get a rough idea on how fast it would be on a modern PC. By "rough idea" I mean within an order of magnitude.

Comparing computer performance over decades it tricky, because in the past we could always make things run much faster by programming the inner loops in microcode. Because of this, PCs 20 years ago were more responsive than a PC today. But then, we progressed from $50,000 ECL PCs consuming 3,000 W to $500 CMOS PCs consuming 15 W. And today's PC use a lot of cycles on fancy animations, sounds, and other GUI effects versus the frugal wabi-sabi (侘寂) GUIs of yore. Finally the system architectures are quite different.

In the old days, computers were rated in MIPS, or million operations per seconds. I thought this should still be a valid measure for comparing the performance in scientific computing. I went to HP's product page to look up the MIPS rating of my PC, but could get no other ratings than CPU clock speed and front side bus speed. Such numbers are not very meaningful, because you do not know how much time the processors are idling waiting for data (very little in the past, a lot today), etc.

I was equally left in the dark on Intel's product page, so I resorted to a search engine. The top result was a Wikipedia page, which rated the CPU in my PC to about 20,000 MIPS. I thought this did not feel right, so I asked the performance specialist on the Performance Agora. Read here what he has to say on modern microprocessor MIPS:

Performance: Close-up of man on wakeboarding reaching down to touch water

On a separate note, if you enjoyed our April's fool post, you might be interested in this serious paper appeared on April 3: H.T. Ng and S. Bose, Entangled light from Bose–Einstein condensates, New J. Phys. 11 (2009) 043009.