A growing body of research attempts to put a number on energy use and AI—even as the companies behind the most popular models keep their carbon emissions a secret.
A problem is that the information is not in the hands of the company selling the AI. The actual hardware is often owned by service providers and independent data centers.
Im sure they can do the simple math of: we pay for x power, we have y customers. x / y would be a rough but probably pretty accurate number if we are talking tens of thousands to millions of customers.
You can produce a remarkably good estimate by looking at CPU and GPU utilization out of procfs and profiling a handful of similar machines power use with similar utilization and workloads.
Network is less than 5% of power use for non-GPU loads; probably less for GPU.
Sure, you can do that at an aggregate level, but then how do you divide it by customer? And even then, some setups will be more efficient than others, so you’d only get that setup’s usage.
And even if you do that and can narrow it down to a single user and a single prompt, you can still only roughly predict how long it will think and how long the response will be.
By customer is easy: they’re each renting specific resources. A fractional cloud instance (excepting the sma burst able ones) is tied to specific CPUs and GPUs. And there are records of who rented which one when being kept already.
You might not be able to break out specific individual queries, but computing averages is completely straightforward
A problem is that the information is not in the hands of the company selling the AI. The actual hardware is often owned by service providers and independent data centers.
They know exactly what the power consumption of that hardware is though. This isnt tough to figure out just because you use a cloud provider
deleted by creator
Im sure they can do the simple math of: we pay for x power, we have y customers. x / y would be a rough but probably pretty accurate number if we are talking tens of thousands to millions of customers.
You can produce a remarkably good estimate by looking at CPU and GPU utilization out of procfs and profiling a handful of similar machines power use with similar utilization and workloads.
Network is less than 5% of power use for non-GPU loads; probably less for GPU.
Sure, you can do that at an aggregate level, but then how do you divide it by customer? And even then, some setups will be more efficient than others, so you’d only get that setup’s usage.
And even if you do that and can narrow it down to a single user and a single prompt, you can still only roughly predict how long it will think and how long the response will be.
By customer is easy: they’re each renting specific resources. A fractional cloud instance (excepting the sma burst able ones) is tied to specific CPUs and GPUs. And there are records of who rented which one when being kept already.
You might not be able to break out specific individual queries, but computing averages is completely straightforward