Hacker News Reader

Comment by briga | Hacker News Reader

On the other hand, as software and frameworks get more complex, they require more compute to run. Software hasn't decoupled economic growth from energy consumption, it has just made the energy consumption harder to see.

Edit: since people are asking for examples, I would have to point to the insane cost of training modern ML models. For instance, according to [1], training a single big NLP model can burn up to 620,000 lbs of CO2.

I say a lot of energy consumption is hidden because it's generally moved to data centres, which by the way use up astronomical amounts of energy [2]

[1] https://arxiv.org/abs/1906.02243

[2] https://ec.europa.eu/jrc/en/publication/trends-data-centre-e...

View on HN

> For instance, according to [1], training a single big NLP model can burn up to 620,000 lbs of CO2.

That's not 'an NLP model', it's one of the largest and baddest NLP models that has to be trained only once and can be reused 1000 times over by fine-tuning on different tasks. The paper was hyperbolising about energy use. The AI labs that can even attempt that kind of things can be counted on the fingers of one hand.

A regular NLP model trains in minutes up to a day on a single GPU, so it uses energy on the level of personal game play. And let's not forget that ML can solve difficult math problems with as little as 1/100 of the energy of classical computing by using smart approximation methods. Neural nets also have a nice degradation effect where you can shrink the net by 20x and lose just a few percentage points of accuracy.

That's why Google's speech recognition and synthesis can now run on a single cell phone instead of a datacenter. And we haven't really started building specialised AI chips that will reduce energy usage even further - we're still using mostly GPUs.

> That's why Google's speech recognition and synthesis can now run on a single cell phone instead of a datacenter.

Can it? I thought it still sent out input to the cloud for processing?

check if it works in airplane mode, it might on new hardware

With junk performance that's not even passable. It's a different model.

I think cryptocurrency takes the cake vs. any sane/non-bug thing that frameworks are probably doing. Maybe some AI workloads are in the ballpark?

AI workloads actually have a business reason to exist though. Bitcoin/cryptocurrency space is still a novelty filled with armchair investors who still think they can get rich from it.

I really doubt any AI compute comes close to the kinds of resources directed at cryptocurrency.

Moore's law has decreased amount of energy required to run the same code.

I think Moore's law was just about compute capability, not energy consumption. I could be wrong, though. That being said, processors have become more efficient. I think the other poster was saying that our code complexity often expands to fill the voids. Obvious examples are neutral nets, render farms, crypto mining. But then there is the cost of React vs vanilla, electron vs native. Efficient processors don't necessarily mean less energy consumption, though I think in general that had still proven to be true.

The way Moore's law worked was that ever more fraction of energy was being used for compute instead of fort waste heat.

I deal with low power microcontrollers. They keep using less and less power while running faster.

Old Z80 based board I first worked on drew 1A at 5V, so 5W. Put your hand on it, it's nice a warm. Current board I'm dicking with draws 8ma at 3V or 0.024W. 200 times less power. But it's 50 to 100 times faster.

Dennard scaling is a pretty fundamental part of Moore's law.

Moore's law is dead, however.

Electron/Chrome undid any gains from Moore's Law. A native IRC client takes negligible amount of energy compared to Slack (I can almost double my laptop's battery life by closing Slack and using weechat/irssi instead).

>Software hasn't decoupled economic growth from energy consumption, it has just made the energy consumption harder to see.

You're going to have to explain that one. We can pretty trivially measure energy usage...

What’s the energy usage of a popular GitHub repo?

Fortnite? Minecraft? AirBnB? Google? Android? iOS?

If it’s trivial, what are the numbers and what’s your confidence in them?

Could you explain this more? I would think energy use is quantifiable when payed for by metered consumption.

It is quantifiable, but it's not easily attributable. That is, you can tell that e.g. New York used up 104 terawatt-hours in one year[0], but it's harder to tell what this energy was used for - and, given telecommunications, how much of it was used to fulfill services around the world that just happen to be using a data center in New York.

Price isn't good approximation either, as it's pretty much arbitrary - prices are determined by trade, fluctuate all the time. Regular people only see a smoothed-out, low frequency version of that.

[0] - 8.6 megapersons living in New York * 12 MWh US average electricity use per person per year.

Maybe there's some misunderstanding here. I'm not saying cost is a good, or even viable, way of quantifying energy consumption.

Since energy suppliers meter their output to apply a price, I'm wondering how hidden costs could exist (unless the software is somehow circumventing or undermining the meters?).

I would think it'd be pretty straight-forward experiment to use software from the late 80s / early 90s for a couple weeks, measure machine energy consumption, then compare with the consumption from a machine running the latest version of everything for two weeks. Next month repeat but also use a processor from the 90s for the first two weeks. Rinse, repeat, extrapolate, apply statistics, &c. Maybe try adjusting for or comparing differences in appliances common then and now.

If software energy costs are somehow hidden, I can only think they're hidden because we're deliberately not choosing to look at them.

It's not, though. You don't pay the electric bill for most services you consume.

Reply on news.ycombinator.com