I’m gonna throw some numbers around based on a conversation with my friend the other day.
He works at an oil refinery nearby and under heavy load the whole facility can use up to 150MW of electricity. That’s while being rated as one of the least environmentally harmful refineries in the US. They have ~300 acres of solar panels supplying ~35MW over the course of a full day (~50MW during daylight hours). That means 20~25% of their power needs are fully renewable.
He also mentioned that the new Meta datacenter going up nearby is projected to use ~400MW! That’s two and a half oil refineries! That also means that to cover its electricity use they would need to set up ~2800 acres of solar panels. That’s ~11 times the footprint of the datacenter itself and two large farms worth of space on some of the best farmland in the country!
They definitely should be 100% powering the facility with renewables, but even then I don’t think model training justifies the space and resources that need to be invested to accomplish that.
So, requiring them to cover their datacenters with solar farms would lessen power requirements by 9% plus whatever offset to cooling costs, that’s not nothing as we try to find ways to power these things beyond private natural gas plants, which they can’t keep doing.
That should definitely be the barest minimum, it’s just insane to me how much one of these data centers demands. Even beyond the privacy/copyright concerns, it doesn’t seem to me like the benefits of training models faster justifies the costs.
The 7.65GW datacenter Amazon wants to make, would need about 155,000 acres to have a 4x provisioned solar array if they wanted to operate it all day off solar based of the US’s largest solar farm. That doesnt account for land space for storage. Also a 4x over provision probably isnt truly enough for something like this to reliably operate off solar only.
Itd be great to make some solar though for daytime hours even if trying to store the ridiculous amount of power required is its own insane problem and not worth it.
I’m gonna throw some numbers around based on a conversation with my friend the other day.
He works at an oil refinery nearby and under heavy load the whole facility can use up to 150MW of electricity. That’s while being rated as one of the least environmentally harmful refineries in the US. They have ~300 acres of solar panels supplying ~35MW over the course of a full day (~50MW during daylight hours). That means 20~25% of their power needs are fully renewable.
He also mentioned that the new Meta datacenter going up nearby is projected to use ~400MW! That’s two and a half oil refineries! That also means that to cover its electricity use they would need to set up ~2800 acres of solar panels. That’s ~11 times the footprint of the datacenter itself and two large farms worth of space on some of the best farmland in the country!
They definitely should be 100% powering the facility with renewables, but even then I don’t think model training justifies the space and resources that need to be invested to accomplish that.
So, requiring them to cover their datacenters with solar farms would lessen power requirements by 9% plus whatever offset to cooling costs, that’s not nothing as we try to find ways to power these things beyond private natural gas plants, which they can’t keep doing.
That should definitely be the barest minimum, it’s just insane to me how much one of these data centers demands. Even beyond the privacy/copyright concerns, it doesn’t seem to me like the benefits of training models faster justifies the costs.
The 7.65GW datacenter Amazon wants to make, would need about 155,000 acres to have a 4x provisioned solar array if they wanted to operate it all day off solar based of the US’s largest solar farm. That doesnt account for land space for storage. Also a 4x over provision probably isnt truly enough for something like this to reliably operate off solar only.
Itd be great to make some solar though for daytime hours even if trying to store the ridiculous amount of power required is its own insane problem and not worth it.