Posted: 03.17.15

What spec works for you?

For as long as I can remember— which is kind of a redundant statement because who can remember longer than they can remember?— we in the computer industry have been quantifying, measuring, and judging the things we make, or use, on the basis of how many times a second they can do something.

We used to talk about MIPS (not the company or processor, but millions of operations per second). Or we spoke about triangles per second; we still do to a certain extent. Over the years, as things kept getting faster, we evolved from one metric to another, but the one that seems to have stuck the longest is FLOPS—FLoating point Operations Per Second—and that’s why the “S” is always capitalized.  FLOPS ARE less useful than these flops.

The first instance I could find of the use of the term FLOPS was to describe the speed of the 1957 IBM 709, which hit 5 kFLOPS. By 1964, the CDC (Control Data Corporation) 6600 (designed by Seymour Cray) hit 500 kFLOPS—a 100x increase in seven years! Almost twice Moore’s law, which hadn’t been derived yet. (By the way, Moore’s law is 50 this year.) Today, Nvidia claims its Tegra X1 can deliver over one TFLOPS—that’s 200 million times faster than the 1957 IBM 709 in just under 70 years, which is less than what Moore’s law should have predicted. However, when you look at China’s supercomputer the NUDT (National University of Defense Technology) Tianhe-2, which can do almost 34 PFLOPS, that’s quite a bit ahead of what Moore’s law would predict.

So what?

Other than being good for amusing bar talk, who cares? And that’s the rub with ROPS, FLOPS, and BOPS—they don’t really tell you anything, other than that one computer or chip may be faster at calculating floating-point math than another. One problem is we don’t even have a truly universally accepted methodology for calculating FLOPS, or ROPS, or almost anything else—it’s subject to, ahem, interpretation.

For example, would a car’s raster-generated dashboard look any better if the GPU driving it did twice the FLOPS? Would you get any better web browsing with your phone if it could produce twice the FLOPS? There are a lot of things we deal with on a daily basis that are good enough, and more FLOPS aren’t going to help them. If that’s the case, then why are we judging those de-vices on the basis of FLOPS?

Because that’s all we know how to do. We’re locked in our engineering bringing-up and can’t seem to bust out of it. We’re becoming the modern-day equivalent of the techno Tower of Babel, everyone talking and no one understanding.
We need a new metric—let’s call it a delight metric. It can’t be based on an arbitrary atomic clock. It has to be based on the wisdom of the crowd. When I copy some text from a web page, and then open up Word and paste it in, I get a blob. Why? Because Microsoft insists you open the app first and then copy and paste—that’s not delightful, that just makes me pissed off at Microsoft, and no amount of FLOPS is going to change Microsoft’s arbitrary user-antagonistic design.

Qualcomm is looking into the next generation of smartphones, which will be able to react to the environment and tell you stuff, warn you, advise you. The devices are going to be delightful, and we won’t be able to measure the delight with FLOPS— but we’ll try if we don’t come up with some means of comparing Qualcomm’s results with Intel’s. Now, there’s the grand challenge: How are we going to judge and evaluate the multitude of devices that will be avail-able to us in the next few years? It can’t be ROPS, FLOPS, or BOPS, but it sure as hell will ultimately be bucks, yuans, and euros.

So maybe the way we measure things going forward is on the PE of the share price. Look at Apple’s and Microsoft’s— hmmm, maybe not such a bad metric after all.