3 ms·

AFAIK it comes down to some degree to a tradeoff between the projected density and power efficiency gains of newer gear vs the cost savings of keeping older gea

AFAIK it comes down to some degree to a tradeoff between the projected density and power efficiency gains of newer gear vs the cost savings of keeping older gear around and in the racks. Some datacenters simply can't support higher density due to power constraints and if the newer gear isn't offsetting the capital costs by pulling in more revenue then you might as well keep the old gear around longer.

jeffbee 5y ago
I don't know why people always mention density. Google isn't constrained on floor space. If they need another 100 acres of Iowa for their servers to stand on, they'll just buy that for a comparative $0 cost. Density is an issue for people who are still renting from Equinix.
- greenail 5y ago
  the substation has a limit, so do all of the power distribution systems.
  - jeffbee 5y ago
    Sure but that's the opposite trend from density, is it not? A dual 64-core EPYC Rome machine - and we know they are using these for web search, from published research papers - has 400W of CPU TDP alone, which is a more than double the per-host TDP of a Westmere EP (as an example of a retired platform). So if anything there is more empty space in the racks in 2021 than there would have been in 2011, based on the design power load of existing buildings.
    - greenail 5y ago
      The cycle for servers is 3 years, the cycle for the building is 30 years. What I'm trying to say is that if you can deliver more widgets per watt per sq ft by replacing the server than you go for that. If you can't why not keep the already depreciated server operating for another year. The incremental costs are mostly the power (and ac) plus the opportunity cost of putting a more efficient server in a year later. If it fails you just put in the newer server. If it doesn't you may save a bit.

3 ms·

AFAIK it comes down to some degree to a tradeoff between the projected density and power efficiency gains of newer gear vs the cost savings of keeping older gea

by greenail 5y ago

jeffbee 5y ago
I don't know why people always mention density. Google isn't constrained on floor space. If they need another 100 acres of Iowa for their servers to stand on, they'll just buy that for a comparative $0 cost. Density is an issue for people who are still renting from Equinix.
- greenail 5y ago
  the substation has a limit, so do all of the power distribution systems.
  - jeffbee 5y ago
    Sure but that's the opposite trend from density, is it not? A dual 64-core EPYC Rome machine - and we know they are using these for web search, from published research papers - has 400W of CPU TDP alone, which is a more than double the per-host TDP of a Westmere EP (as an example of a retired platform). So if anything there is more empty space in the racks in 2021 than there would have been in 2011, based on the design power load of existing buildings.
    - greenail 5y ago
      The cycle for servers is 3 years, the cycle for the building is 30 years. What I'm trying to say is that if you can deliver more widgets per watt per sq ft by replacing the server than you go for that. If you can't why not keep the already depreciated server operating for another year. The incremental costs are mostly the power (and ac) plus the opportunity cost of putting a more efficient server in a year later. If it fails you just put in the newer server. If it doesn't you may save a bit.