The Math Behind GPU Power Consumption And PSUs

Garbage in, garbage out. That's what you get if you don't spend enough time planning and testing your benchmarking methodologies. It's easy to fall victim to this golden rule, since preparation requires a lot of effort, eating into nights and weekends. It has taken me a year to find a solution that I’m really happy with. I’ve gained a lot of insights along the way, and discovered some phenomenon I really didn’t expect.

I originally set out to improve our power consumption measurements for graphics cards. I felt that the usual way of doing this via estimates based on measurements at the power plug just didn’t cut it, and I wanted a way to get more reliable and valid results with some extra effort.

Simple current clamps, multimeters and the first data loggers never quite worked for me. Something was always off, and it got really frustrating when every measurement arrived at a different result. There was some factor I was just not taking into account. But what could it be?

That's exactly what I want to diagnose over the next few pages. Unfortunately, I can’t spare you a certain minimum amount of theory. It’s exactly this knowledge that’s so important if you want to understand why the simple solutions don’t work. Never fear, though. The story won’t be boring or dry, and I’ve boiled down and simplified the subject matter as much as I could.

The measurement setup was systematically changed and improved time and time again in cooperation with our industrial partner. Entire configurations were tested, only to be scrapped again. Firmware was updated, nearly putting us back to square one. All in all, creating this platform was a lot of work on top of my already time-sensitive assignments, especially since everything had to be built in the midst of our normal operation.

All of the effort was really worthwhile, though. And, when it really comes down to it, transparency always beats secrecy. The latter might be good for business, but, in the end, you need to be able to back up your claims, especially if you publish them. This is why I’ll tell you the secret of how we get to the results produced by our measurements. Otherwise, you might just not believe them…