Posted on Sat 13 December 2014

lies, damned lies and benchmarks

Rule of thumb: if you have a working computer, you shouldn't replace it with one that's less than twice as fast.

If Moore's Law held true, then that means you can spend the same amount every 24 months and double your speed. Turns out, Moore's Law isn't true. Kinda. Sorta.

OK, so Moore's Law applies to chip density, not megahertz or any other measure of performance, and it's an observation, not a real law. There are periods of fluctuation, and the point at which it is economical to produce CPUs is always 1-3 years behind the design phase. Right now commercial feature size is around 22 nanometers; if you could believe Intel and IBM and AMD in 2010, 14 nm fabs should have started producing commercial chips this year. They haven't.

So let's go to a benchmark: the PassMark Single Thread CPU Benchmark. See all those caps? That means they won't tell you quite what they're doing, but you can compare scores. Not knowing what PassMark is actually doing is problematic for comparing things that are very different from each other, but here we are going to end up being within an order of magnitude from top to bottom anyway.

Here are some sample CPUs, and their reported scores, number of cores, and a naive extrapolation (approximately multiplying cores by single-thread score, to represent a sort of ideal parallelized performance):

CPU Score #cores S*n
AMD 5350 804 4 3200
AMD Athlon X4 640 878 4 3500
Intel Core 2 E6600 920 2 1800
AMD FX-4130 1269 4 5000
Intel Core i3-3225 1786 2 3600
Intel Core i5-2500 1896 4 7600
Intel Pentium G3258 2179 2 4300
Intel Core i3-4770 2240 2 4500

You may notice that all of these are cheap-ish desktop CPUs. I don't play games on these machines, and I especially don't play 3D games. The 5350 is in fact marketed as one step up from an embedded system (and I'm using it in my new firewall); the i3 series is supposed to be Intel's basic business desktop. Not one of them has ever been the top-of-the-line.

So, in everyday use, does either the score or the score*#cpus represent how fast the machines feel? There's a factor of 3, almost, in single thread scores, and a factor of 6 in the parallel scores.

Kinda, sorta, not really.

I don't run complex calculations for the sake of doing the math. I push a lot of pixels, but only in 2 dimensions, which barely counts these days. And the truth is, I don't really feel the difference in CPU speed between the E6600 and the i5, both of which I use as desktops on a daily basis.

In fact, I don't really notice any CPU differences. Here's what I notice:

  • the difference between machines with just enough RAM and loads of RAM
  • the difference between machines with SSDs and spinning disks

That's it. Want a fast machine for cheap? Take a low-end CPU, add lots of RAM (16GB and up) and get a reliable SSD. Back it up automatically every night. You can build it yourself for about $450 or so, maybe less.

For what it's worth, PassMark says that the best single-thread cpu on the planet tops out at 2535, which is a little more than 3x the performance of chips so cheap you can barely buy them anymore. And remember that the Pentium G3258, at 2179, is only $60 or so. That looks pretty promising.

© -dsr-. Send feedback or comments via email — by continuing to use this site you agree to certain terms and conditions.

Built using Pelican. Derived from the svbhack theme by Giulio Fidente on github.