Sounds like we need to scale down generative AI. Honestly.. I don't buy it. There's this fundamental idea being sold here that we'll just scale up electrical grids till infinity, but in reality...
Sounds like we need to scale down generative AI.
Honestly.. I don't buy it. There's this fundamental idea being sold here that we'll just scale up electrical grids till infinity, but in reality that's just not an option. "Can't handle the load" is implying it should and maybe there are more important things we should scale up for instead.
Comment box Scope: comment response, personal opinion Tone: neutral Opinion: yes Sarcasm/humor: none IDK. I use ChatGPT at work and it helps me. I think it helps a lot of people. Instead of trying...
Comment box
Scope: comment response, personal opinion
Tone: neutral
Opinion: yes
Sarcasm/humor: none
IDK. I use ChatGPT at work and it helps me. I think it helps a lot of people.
Instead of trying to 'scale down generative AI' maybe a better conceptualization would be to configure the market such that heavy electricity users actually pay for what maintenance and expansion costs they incur??? From what I know, economists agree that economies work when people pay for the services they use.
People have been doing increasingly more computationally intensive and energy-inefficient tasks with computers and other electricity-based machines for decades. Last year this same kind of article was being written about crypto miners. I'm sure next year it will be some other energy-intensive fad. I don't think regulating based on specific use-cases scales well.
In contrast, it would be logistically trivial for the federal government to levy an additional tax on any electricity user using more than X amount of electricity per some time period, and have this amount increase by O(logn) or O(n^2) or something for every additional unit per time period. If small electricity users (like households) are omitted, then this tax would only impact large businesses like datacenters.
An optimization would be to make this additional fee vary in real time, based on how much strain the grid is under. The grid is definitely not at capacity 100% of the time. Lots of people already have this kind of payment plan for their electricity. It could just be more heavily incentivized for large electricity users.
This would discourage such enterprises from using that much electricity in the exact way they are doing so:
Some large electricity users would stop running certain energy-intensive tasks at certain times.
Some large electricity users would find ways to be more energy-efficient to avoid spending so much.
Some large electricity users would go off-grid and pay for their own infrastructure and electricity generation/transmission. In this economy their solution would almost definitely be solar + batteries.
The solution isn't different than any current fees electricity users pay, except that it would be a little higher for people using a lot of electricity.
The end result is that the government would be have less of a deficit on maintenance for a given output product which theoretically benefits society, and conceivably would have less of a deficit in absolute terms. If a company decides to go off-grid with its gen AI, they can fund that however they like, such as by charging users more for the services they're using.
You don't even need a tax for this purpose, the power companies themselves could just charge whatever they need to in order to expand their grid. I don't know how exactly it's done in the US, but...
You don't even need a tax for this purpose, the power companies themselves could just charge whatever they need to in order to expand their grid.
I don't know how exactly it's done in the US, but here in Germany, for large electricity consumers (typically industrial, but not necessarily) their prices for each kWh of power gets determined by the peak consumption rate within the month, smoothed over 15 minute intervals. So it's not just how much power you consumed or the average, but your highest rate of consumption. And this can result in significant differences in electricity bills, so much so that considerable effort goes into power monitoring and management to ensure the peak consumption rate is as low as possible.
If US power companies would implement something similar for data centers, they could potentially raise the funds needed.
There's many different power companies, some might do that, but generally they simply charge by the kWh. Mine does send a message saying that peak times are coming up and if I have reduced usage I...
There's many different power companies, some might do that, but generally they simply charge by the kWh.
Mine does send a message saying that peak times are coming up and if I have reduced usage I may get a rebate. I don't know how they actually check that, I'm guessing by distro circuit rather than individual meters based on how quickly they update us. The rebate is generally in the range of a couple cents to a couple bucks, virtually worthless.
I work with power line systems daily (when I'm not slacking off), my current project area is mentioned in the article. I fear for the reliability of our aging grid considering the lack of...
I work with power line systems daily (when I'm not slacking off), my current project area is mentioned in the article. I fear for the reliability of our aging grid considering the lack of preventative maintenance, tech to identify potential issues, and funding. I pretty regularly see cable that hasn't been updated in over 100yrs. You can't just slap a sensor on every transformer, that's just not feasible for one (a city the size of Detroit will have tens of thousands of miles of cable just for distro), and two, it doesn't capture the mechanical failures from rusting parts, pole lean, damage to wood, etc.
We could iginite a New New Deal just from checking out and updating our power grid, we need that more than more gen AI or Amazon warehouses. Lives literally depend on it. Such a deal could be super-charged by pushing for public Internet and fiber installation as well, but that's a whole other tangent.
Kind of sounds like the main problem is utilities not charging data centers enough for their resource needs. Hard to find enough power for their needs? Charge them more to mediate demand. Need to...
Kind of sounds like the main problem is utilities not charging data centers enough for their resource needs. Hard to find enough power for their needs? Charge them more to mediate demand. Need to expand power lines and replace transformers? Charge them more to cover the upgrade costs. Not enough water for cooling? Charge more for water. Keep increasing the price until the demand and supply balance out. I kind of wonder if that's what is happening already and these articles are just here to make us feel more worried than we should be.
Comment box Scope: providing information Tone: neutral Opinion: none Sarcasm/humor: none Solar power will supply half of the increase in demand for the world next year. Not half of the world's...
Comment box
Scope: providing information
Tone: neutral
Opinion: none
Sarcasm/humor: none
Solar power will supply half of the increase in demand for the world next year. Not half of the world's power needs. These metrics are distinct.
As I understand it, it is two different problems. Even if we could generate all the power needed, the grid will not necessarily be able to distribute the increased demand. There is also the...
As I understand it, it is two different problems. Even if we could generate all the power needed, the grid will not necessarily be able to distribute the increased demand. There is also the concern that the power needed for AI is greater than the increase in green energy output, so fossil energy is also increasing. Google and other tech companies are struggling to meet their own green energy ambitions.
Whenever I read about some technology and the current power grid system "not being able to handle it", no matter where I read the headline, I rarely see a critique against the power companies for...
Whenever I read about some technology and the current power grid system "not being able to handle it", no matter where I read the headline, I rarely see a critique against the power companies for not maintaining the grid or being forward thinking enough to prepare for power demand. While these newer technologies consume a lot of power why does this keep becoming the narrative and not "does the privatized power industry deter innovation due to the requirement of profits over consumption needs?" It was crypto 5 years ago and now generative AI... What's the next new technology that "can't work" because its power needs are greater than what privatized power companies want to handle? Nationalize utilities if power companies refuse to spend on needed operating expenses versus "generating shareholder value".
Datacenters in general, and GPU workloads in particular, are a perfect example of a relatively centralised resource that doesn’t need to be physically near the customer (mostly). I think the...
Datacenters in general, and GPU workloads in particular, are a perfect example of a relatively centralised resource that doesn’t need to be physically near the customer (mostly). I think the comments above pointing out that the utilities should be upgrading the grid are quite right, but when you’ve got a massive power sink in one place it’s also a solid candidate for a dedicated solar or wind farm plugged right into the side to take some of the overall load.
Sounds like we need to scale down generative AI.
Honestly.. I don't buy it. There's this fundamental idea being sold here that we'll just scale up electrical grids till infinity, but in reality that's just not an option. "Can't handle the load" is implying it should and maybe there are more important things we should scale up for instead.
Comment box
IDK. I use ChatGPT at work and it helps me. I think it helps a lot of people.
Instead of trying to 'scale down generative AI' maybe a better conceptualization would be to configure the market such that heavy electricity users actually pay for what maintenance and expansion costs they incur??? From what I know, economists agree that economies work when people pay for the services they use.
People have been doing increasingly more computationally intensive and energy-inefficient tasks with computers and other electricity-based machines for decades. Last year this same kind of article was being written about crypto miners. I'm sure next year it will be some other energy-intensive fad. I don't think regulating based on specific use-cases scales well.
In contrast, it would be logistically trivial for the federal government to levy an additional tax on any electricity user using more than X amount of electricity per some time period, and have this amount increase by
O(logn)orO(n^2)or something for every additional unit per time period. If small electricity users (like households) are omitted, then this tax would only impact large businesses like datacenters.An optimization would be to make this additional fee vary in real time, based on how much strain the grid is under. The grid is definitely not at capacity 100% of the time. Lots of people already have this kind of payment plan for their electricity. It could just be more heavily incentivized for large electricity users.
This would discourage such enterprises from using that much electricity in the exact way they are doing so:
The solution isn't different than any current fees electricity users pay, except that it would be a little higher for people using a lot of electricity.
The end result is that the government would be have less of a deficit on maintenance for a given output product which theoretically benefits society, and conceivably would have less of a deficit in absolute terms. If a company decides to go off-grid with its gen AI, they can fund that however they like, such as by charging users more for the services they're using.
You don't even need a tax for this purpose, the power companies themselves could just charge whatever they need to in order to expand their grid.
I don't know how exactly it's done in the US, but here in Germany, for large electricity consumers (typically industrial, but not necessarily) their prices for each kWh of power gets determined by the peak consumption rate within the month, smoothed over 15 minute intervals. So it's not just how much power you consumed or the average, but your highest rate of consumption. And this can result in significant differences in electricity bills, so much so that considerable effort goes into power monitoring and management to ensure the peak consumption rate is as low as possible.
If US power companies would implement something similar for data centers, they could potentially raise the funds needed.
There's many different power companies, some might do that, but generally they simply charge by the kWh.
Mine does send a message saying that peak times are coming up and if I have reduced usage I may get a rebate. I don't know how they actually check that, I'm guessing by distro circuit rather than individual meters based on how quickly they update us. The rebate is generally in the range of a couple cents to a couple bucks, virtually worthless.
I work with power line systems daily (when I'm not slacking off), my current project area is mentioned in the article. I fear for the reliability of our aging grid considering the lack of preventative maintenance, tech to identify potential issues, and funding. I pretty regularly see cable that hasn't been updated in over 100yrs. You can't just slap a sensor on every transformer, that's just not feasible for one (a city the size of Detroit will have tens of thousands of miles of cable just for distro), and two, it doesn't capture the mechanical failures from rusting parts, pole lean, damage to wood, etc.
We could iginite a New New Deal just from checking out and updating our power grid, we need that more than more gen AI or Amazon warehouses. Lives literally depend on it. Such a deal could be super-charged by pushing for public Internet and fiber installation as well, but that's a whole other tangent.
Kind of sounds like the main problem is utilities not charging data centers enough for their resource needs. Hard to find enough power for their needs? Charge them more to mediate demand. Need to expand power lines and replace transformers? Charge them more to cover the upgrade costs. Not enough water for cooling? Charge more for water. Keep increasing the price until the demand and supply balance out. I kind of wonder if that's what is happening already and these articles are just here to make us feel more worried than we should be.
This is weird in contrast with the other headline that solar power is going to supply half the power needs for the world next year
Comment box
Solar power will supply half of the increase in demand for the world next year. Not half of the world's power needs. These metrics are distinct.
As I understand it, it is two different problems. Even if we could generate all the power needed, the grid will not necessarily be able to distribute the increased demand. There is also the concern that the power needed for AI is greater than the increase in green energy output, so fossil energy is also increasing. Google and other tech companies are struggling to meet their own green energy ambitions.
Oh I'm not trying to say it isn't a problem, I just found the comparison of these two headlines humorous is all.
Whenever I read about some technology and the current power grid system "not being able to handle it", no matter where I read the headline, I rarely see a critique against the power companies for not maintaining the grid or being forward thinking enough to prepare for power demand. While these newer technologies consume a lot of power why does this keep becoming the narrative and not "does the privatized power industry deter innovation due to the requirement of profits over consumption needs?" It was crypto 5 years ago and now generative AI... What's the next new technology that "can't work" because its power needs are greater than what privatized power companies want to handle? Nationalize utilities if power companies refuse to spend on needed operating expenses versus "generating shareholder value".
Datacenters in general, and GPU workloads in particular, are a perfect example of a relatively centralised resource that doesn’t need to be physically near the customer (mostly). I think the comments above pointing out that the utilities should be upgrading the grid are quite right, but when you’ve got a massive power sink in one place it’s also a solid candidate for a dedicated solar or wind farm plugged right into the side to take some of the overall load.