There is a correlation between the training time and energy consumption, but that doesn’t mean there is a correlation between training time and carbon emissions.
Thanks for your article. I am going to annoy you one more time, I hope that's ok :)
1. You seem to be focusing your articles on electricity consumption. That is only a tiny part of the carbone footprint of the digital industry. For instance, you mention that new HW is more efficient. That implies acquiring new HW, which implies their manufacture (and transport). Manufacture of IT HW is very greedy in term of minerals and energy, which is actually most of the carbon footprint. What kind of energy mix you use to power that has little influence on the total footprint.
2. As you point out, the fact that HW and algorithms are becoming more efficient has absolutely no impact on the amount of electricity Google uses to power his AI (and other services). You can reduce by 2 the power needed to run a service, if you multiply by 3 the amount of users or services, you end-up having larger carbon emissions in total (which is what we are concerned about).
3. Now, regarding the energy mix, you need to take into account that Google is not the only user of the power grid in periods of high wind and sunny days... When google schedule its heavy computations during periods of peak of decarbonized energy, it prevents other users from using this mix. This is because the amount of 'green' energy is finite. It is actually so finite that 100% produced is consumed at all times. So, assuming Google could use exclusively decarbonized energy, it would simply push other users towards carbonized sources, but the global picture would not change: the same amount of carbon is emitted at the scale of the country, while only Google's carbon footprint would appear "greener" as an artefact.
Thanks for your comment. As you rightly point out, energy & carbon are just one aspect of the overall environmental impact of computing. I've written about water quite a bit, for example. It's also true that for consumer products, the manufacturing stage is the largest in terms of the carbon footprint.
However, in the data center it is the opposite. The use stage is the largest component of the carbon footprint, not the hardware manufacturing. I wrote about that a few years ago - https://davidmytton.blog/carbon-footprint-laptops-vs-servers-intel-vs-arm/ - and you can see it in the latest LCA reports from manufacturers such as Dell. AI training is generally done in data centers, at least at large scale, which is why I focused on that here.
As you pointed out in your previous article, Dell calculations are non-standard, and in general, I would not trust the carbon footprint provided by the manufacturer until carbon footprints are imposed by law and therefore double checked by independent organisms (as it starts to be the case in Europe). That being said, if we trust Dell blindly, the manufacture of 1 server is 1 ton of CO2eq. Not really insignificant, is it?
In addition if I quote when you are presenting results for the Dell R930:
"85% of that is in the manufacturing but the variance is so large that this breakdown can’t be trusted."
Here the report tells you that most of the emissions comes from the manufacture, but you seem to decide rather arbitrarily (?) to disregard this result.
But your answer above only addresses pt1. I am curious if you have any feedback about pt3 (honestly, I might have overlooked something!).
Thanks for your article. I am going to annoy you one more time, I hope that's ok :)
1. You seem to be focusing your articles on electricity consumption. That is only a tiny part of the carbone footprint of the digital industry. For instance, you mention that new HW is more efficient. That implies acquiring new HW, which implies their manufacture (and transport). Manufacture of IT HW is very greedy in term of minerals and energy, which is actually most of the carbon footprint. What kind of energy mix you use to power that has little influence on the total footprint.
2. As you point out, the fact that HW and algorithms are becoming more efficient has absolutely no impact on the amount of electricity Google uses to power his AI (and other services). You can reduce by 2 the power needed to run a service, if you multiply by 3 the amount of users or services, you end-up having larger carbon emissions in total (which is what we are concerned about).
3. Now, regarding the energy mix, you need to take into account that Google is not the only user of the power grid in periods of high wind and sunny days... When google schedule its heavy computations during periods of peak of decarbonized energy, it prevents other users from using this mix. This is because the amount of 'green' energy is finite. It is actually so finite that 100% produced is consumed at all times. So, assuming Google could use exclusively decarbonized energy, it would simply push other users towards carbonized sources, but the global picture would not change: the same amount of carbon is emitted at the scale of the country, while only Google's carbon footprint would appear "greener" as an artefact.
Thanks for your comment. As you rightly point out, energy & carbon are just one aspect of the overall environmental impact of computing. I've written about water quite a bit, for example. It's also true that for consumer products, the manufacturing stage is the largest in terms of the carbon footprint.
However, in the data center it is the opposite. The use stage is the largest component of the carbon footprint, not the hardware manufacturing. I wrote about that a few years ago - https://davidmytton.blog/carbon-footprint-laptops-vs-servers-intel-vs-arm/ - and you can see it in the latest LCA reports from manufacturers such as Dell. AI training is generally done in data centers, at least at large scale, which is why I focused on that here.
As you pointed out in your previous article, Dell calculations are non-standard, and in general, I would not trust the carbon footprint provided by the manufacturer until carbon footprints are imposed by law and therefore double checked by independent organisms (as it starts to be the case in Europe). That being said, if we trust Dell blindly, the manufacture of 1 server is 1 ton of CO2eq. Not really insignificant, is it?
In addition if I quote when you are presenting results for the Dell R930:
"85% of that is in the manufacturing but the variance is so large that this breakdown can’t be trusted."
Here the report tells you that most of the emissions comes from the manufacture, but you seem to decide rather arbitrarily (?) to disregard this result.
But your answer above only addresses pt1. I am curious if you have any feedback about pt3 (honestly, I might have overlooked something!).