Technology

Up until the early 2000s, serial computation speed doubled about every 18 months. That meant that virtually all software just ran twice as quickly every 18 months of CPU advances. And since taking advantage of that was trivial, new software releases did, traded CPU cycles for shorter development time or more functionality, demanded current hardware to run at a reasonable clip.

In that environment, it was quite important to upgrade the CPU.

But that hasn't been happening for about twenty years now. Serial computation speed still increases, but not nearly as quickly any more.

This is about ten years old now:

https://preshing.com/20120208/a-look-back-at-single-threaded-cpu-performance/

Throughout the 80’s and 90’s, CPUs were able to run virtually any kind of software twice as fast every 18-20 months. The rate of change was incredible. Your 486SX-16 was almost obsolete by the time you got it through the door. But eventually, at some point in the mid-2000’s, progress slowed down considerably for single-threaded software – which was most software.

Perhaps the turning point came in May 2004, when Intel canceled its latest single-core development effort to focus on multicore designs. Later that year, Herb Sutter wrote his now-famous article, The Free Lunch Is Over. Not all software will run remarkably faster year-over-year anymore, he warned us. Concurrent software would continue its meteoric rise, but single-threaded software was about to get left in the dust.

If you’re willing to trust this line, it seems that in the eight years since January 2004, mainstream performance has increased by a factor of about 4.6x, which works out to 21% per year. Compare that to the 28x increase between 1996 and 2004! Things have really slowed down.

We can also look at about the twelve years since then, which is even slower:

https://www.cpubenchmark.net/compare/2026vs6296/Intel-i7-4960X-vs-Intel-Ultra-9-285K

This is using a benchmark to compare the single-threaded performance of the i7 4960X (Intel's high-end processor back at the start of 2013) to that of the Intel Ultra 9 285K, the current one. In those ~12 years, the latest processor has managed to get single-threaded performance about (5068/2070)=~2.448 times the 12-year-old processor. That's (5068/2070)^(1/12)=1.07747, about a 7.7% performance improvement per year. The age of a processor doesn't matter nearly as much in that environment.

We still have had significant parallel computation increases. GPUs in particular have gotten considerably more powerful. But unlike with serial compute, parallel compute isn't a "free" performance improvement -- software needs to be rewritten to take advantage of that, it's often hard to parallelize solving problems, and some problems cannot be solved in parallel.

Honestly, I'd say that the most-noticeable shift is away from rotational drives to SSDs -- there are tasks for which SSDs can greatly outperform rotational drives.