Replies: 10 comments
-
This is fantastic @jawache - thank you for putting this together. We definitely have something concrete to walk through. I know you're not going to be able to make the meeting today so I'd love to save this for next week's meeting when you're back so that you can answer questions around this as we discuss this. @buchananwp something like this would be ideal for discussing from an AzureML perspective. @dtoakley from your work at TW, are there other case studies that you could suggest that will help raise some of the questions like @jawache has towards the end? |
Beta Was this translation helpful? Give feedback.
-
Scott Roscoff from Windows is already doing work in this space, and has agreed to join us. Sending an email to @seanmcilroy29 @dtoakley @atg-abhishek |
Beta Was this translation helpful? Give feedback.
-
Fantastic @buchananwp thank you for making it happen! @jawache should we then defer discussion of this issue till Scott is able to join the call as well? (@buchananwp do we know if he'll be joining today or next week?) |
Beta Was this translation helpful? Give feedback.
-
Thanks @buchananwp would be interested to get his input. I like the idea of deciding on our initial set of characteristics and then running through a few different uses cases, the characteristics are the guard rails in the brainstorming, make sure we don't go into a direction that breaks one of our characteristics.
|
Beta Was this translation helpful? Give feedback.
-
Might be interesting to add DAPR or kind of multi-cloud abstraction - I can see those layers adding overhead, but also an agnostic framework for instrumenting for telemetry. |
Beta Was this translation helpful? Give feedback.
-
Asim: Some things might not include hardware, in some places hardware might not be considered at all. Abhishek: The actions taken by the people designing TF would be different from those who are using TF as a service. Will: it’s a taxonomy of decisions, each with a different sphere of influence Sara: you also want the developers who are using the services to use them smartly where you might have made the API to be green but if people are using it in a way that makes a lot of calls then that negates that green gain Asim: what we need to think about is if taking the action reduces the carbon emissions Vaughn: if we propose numbers that are software-driven then that might be hard for people to act in; there might be different numbers for different parts of the software solution Henry: if we're measuring something and we can't take an action against it, then there is no point of the measurement either Dan: not being able to take an action might be a political thing Asim: software, architecture, or any other choice. actions fall into : (1) use less energy (2) use less hardware (3) using the energy more intelligently (carbon-aware) Vaughn: coming up with a number (an index) - lower number is better would be the good way to go ; it needs to be measured in a consistent way Asim: the number doesn't need to be accurate, as long as it is directional that might be enough |
Beta Was this translation helpful? Give feedback.
-
Taylor: people understand the negative consequences of 'emissions' more than 'kWh' |
Beta Was this translation helpful? Give feedback.
-
Also mentioned - we may want to talk about the actions that are possible; and potentially support those in a future version - but to my mind, anything measured as a KPI still drives change. What is measured drives behaviour change. That’s not to say actions shouldn’t be recommended, but I think the measuring in and of itself is useful - and still has a point. |
Beta Was this translation helpful? Give feedback.
-
What is the next step that we want to take here @jawache ? Should we continue to seek some more details here and then revisit this when we reach a bit more concreteness in the SCI? I think the discussions here have already helped to shape a bit the SCI and the direction that it is taking so this has been very useful!! |
Beta Was this translation helpful? Give feedback.
-
I'm going to propose in today's meeting that we create a rubric for evaluating submissions. I don't think we need to go overboard just a selection of uniquely different application domains. Just to keep us honest, something that works for the ML world might make no sense in the Windows world, the Web world. Let's just run through a list together and use it as an evaluation tool. Later on I think we'll need to work through some rough examples for each application domain, there is space in the spec at the bottom to put that stuff. |
Beta Was this translation helpful? Give feedback.
-
Getting a little bit more concrete here, I think Windows is an excellent use case to explore. Windows really shows the weakness in the GHG attributional model of emissions calculation.
In the GHG attributional model, the carbon emissions of Windows for Microsoft are JUST the carbon emissions from running Windows at Microsoft. I.e. we only count the electricity we bought to run Windows for Microsoft employees. However, Windows is run on 1 billion devices worldwide. Each individual company that uses Windows calculates its own carbon emissions from its own use of Windows. If we add them all up together we get the carbon emissions of Windows for the whole world.
But because the metric we use is the GHG attributional total for Microsoft, the Windows team is not as incentivized as it could be. If they did some work to make WIndows more energy-efficient it would benefit the whole world and every company that uses Windows, but it would only count a little bit towards Microsoft meeting its own carbon emissions targets.
In the SCI contributional model, Windows would calculate the total carbon emissions of all of Windows, worldwide, regardless of which organization is using it and who bought the electricity to run Windows. That way the Windows team is incentivized to do a lot more to improve the energy efficiency of Windows.
NOTE: The Windows team at MSFT doesn't only use the GHG attributional model to measure its carbon emissions for this very reason.
Getting to specifics.
The carbon cost of Windows = Global Energy Usage * Carbon Intensity + Embodied Carbon
Global Energy Usage
Carbon Intensity
Embodied Carbon
Interesting thoughts:
Beta Was this translation helpful? Give feedback.
All reactions