Comment by dahart | Hacker News Reader

Comment by dahart | Hacker News Reader

> A faster language is not always the most energy efficient

My rule of thumb now is that memory access, not compute, is the primary consumer of energy. That would tend to confirm the above statement while also supporting the data that scripting languages aren’t super energy efficient, since they tend to do a lot more dynamic allocation than compiled languages, when generally broadly speaking about common programming practices in each language.

This is mainly colored by GPU usage and a paper/presentation some friends made: https://www.researchgate.net/publication/324217073_A_Detaile...

In the GPU case, memory access costs sometimes 10x more than compute, meaning that minimizing average (not peak) memory traffic is more or less the only path to significantly reduced energy. (See figs 7 & 10 in the linked paper)

"Costs more" in terms of energy per... what quantity exactly? Instruction executed? Clock cycle? Datum processed?

In the paper it’s energy per rendered pixel. You could translate that abstractly to watts per flop.

I think practically, for useful analysis in business it would be wattage per unit of useful work (or datum processed as you suggest). For instance, I work in finance, so what is the wattage per option price calculation?

The reason I asked was that it wasn't clear to me if the parent comment was measuring the same way or not re: memory accesses. It seemed pretty expected that a memory access would be more expensive per byte (it takes like hundreds of clock cycles...) but it wasn't obvious to me if it would be so per clock so I wanted to clarify which was meant.

I absolutely agree that it wasn't clear. I don't think wattage per clock cycle is a meaningful measure here - after all, all languages will have the same wattage per clock cycle, if they're executing the same instructions.

What's important is wattage per unit of useful end product, in my example pricing an option. For others, it might be wattage per web page served or anything else.

Energy per operation and delay per operation, on the order of ~40-100X more than compute...

This is probably true, but it would be incredibly cool to see a research project that demonstrates this.

Reply on news.ycombinator.com