[...] It's an interesting idea, and at least superficially makes sense to me. Commercial buildings save on the capital and maintenance costs of a normal boiler/furnace/heat pump/whatever, data...
The heat generated by a washing-machine-sized data centre is being used to heat a Devon public swimming pool.
The computers inside the white box are surrounded by oil to capture the heat - enough to heat the pool to about 30C 60% of the time, saving Exmouth Leisure Centre thousands of pounds.
The data centre is provided to the council-run centre free of charge.
Start-up Deep Green charges clients to use its computing power for artificial intelligence and machine learning.
Founder Mark Bjornsgaard said the company would also refund the leisure centre's electricity costs for running the "digital boiler" - and seven other England pools had signed up to the scheme.
[...]
"Data centres have got a huge problem with heat," Mr Bjornsgaard said.
"A lot of the money that it costs to run a data centre is taken up in getting rid of the heat.
"And so what we've done is taken a very small bit of a data centre to where the heat is useful and required."
It's an interesting idea, and at least superficially makes sense to me. Commercial buildings save on the capital and maintenance costs of a normal boiler/furnace/heat pump/whatever, data company gets free cooling, physical space, and geographic redundancy. Environmentally it seems a no brainer to use heat where heat is needed rather than spending even more energy on cooling to draw it away.
Right now the data company are paying the energy costs in full, but if it catches on I'd expect that to slide to a split cost model where both still end up paying a bit less than they would if they took on their own bills separately.
Immediate challenges I see are around security, reliability, and maybe economies of scale. Physical access controls on these things aren't going to be anything like up to major server room standards, so that's going to need to be taken into account with whatever data you're sending to them. Connectivity to push that data is also a question, even if compute nodes are less sensitive to it than other types of server - who's paying for that, and can a (presumably) standard gigabit line to the outside world keep this many GPUs fed efficiently?
The chances of losing power or network are going to be way higher than in a multiply redundant centralised setting, and while setting up your software to tolerate the loss of half a rack when someone pulls the wrong breaker shouldn't be too hard, a regional ISP outage that hits multiple locations could be a killer here. SLAs won't be the same so you could be waiting a lot longer for things to come back up. I also wonder if just applying the same approach centrally could be more efficient? I know some datacenters already make good use of their waste heat, although I have no background on how widespread or cost effective that is.
The main issue I'd find here is that compute needs and heating needs don't actually have any relationship to each other. So how would you ensure that you're getting enough heat when you need it...
The main issue I'd find here is that compute needs and heating needs don't actually have any relationship to each other. So how would you ensure that you're getting enough heat when you need it and not overproducing when you don't?
I suppose a swimming pool is a good use-case, since water has such a high specific heat and functionally acts as a giant "heat battery" on its own. But otherwise I think you'd need quite a bit of thermal mass around the data center just to "load balance."
Indeed - I was just talking with someone who was evaluating a new 4-node server. They measured the power input at 250W x 4, so it is truly almost a 1.1kW space heater. The amount of heat to reject...
Indeed - I was just talking with someone who was evaluating a new 4-node server. They measured the power input at 250W x 4, so it is truly almost a 1.1kW space heater. The amount of heat to reject is significant and only increasing.
[...]
It's an interesting idea, and at least superficially makes sense to me. Commercial buildings save on the capital and maintenance costs of a normal boiler/furnace/heat pump/whatever, data company gets free cooling, physical space, and geographic redundancy. Environmentally it seems a no brainer to use heat where heat is needed rather than spending even more energy on cooling to draw it away.
Right now the data company are paying the energy costs in full, but if it catches on I'd expect that to slide to a split cost model where both still end up paying a bit less than they would if they took on their own bills separately.
Immediate challenges I see are around security, reliability, and maybe economies of scale. Physical access controls on these things aren't going to be anything like up to major server room standards, so that's going to need to be taken into account with whatever data you're sending to them. Connectivity to push that data is also a question, even if compute nodes are less sensitive to it than other types of server - who's paying for that, and can a (presumably) standard gigabit line to the outside world keep this many GPUs fed efficiently?
The chances of losing power or network are going to be way higher than in a multiply redundant centralised setting, and while setting up your software to tolerate the loss of half a rack when someone pulls the wrong breaker shouldn't be too hard, a regional ISP outage that hits multiple locations could be a killer here. SLAs won't be the same so you could be waiting a lot longer for things to come back up. I also wonder if just applying the same approach centrally could be more efficient? I know some datacenters already make good use of their waste heat, although I have no background on how widespread or cost effective that is.
The main issue I'd find here is that compute needs and heating needs don't actually have any relationship to each other. So how would you ensure that you're getting enough heat when you need it and not overproducing when you don't?
I suppose a swimming pool is a good use-case, since water has such a high specific heat and functionally acts as a giant "heat battery" on its own. But otherwise I think you'd need quite a bit of thermal mass around the data center just to "load balance."
You'd be shocked at how efficient computers are at generating heat. Gaming PCs are genuinely as efficient as space heaters.
Indeed - I was just talking with someone who was evaluating a new 4-node server. They measured the power input at 250W x 4, so it is truly almost a 1.1kW space heater. The amount of heat to reject is significant and only increasing.