I attended a workshop on ”Sustainable Scientific Computing” at the Leiden Center. It was fantastic, I don’t think I’ve ever attended such a gathering of nerdy solarpunkers. So here is a long braindump in one thread. I have time, because I’m on a 14 hour train trip back home. So far, I’m on time, only 4.5 hours left.
https://www.lorentzcenter.nl/sustainable-scientific-computing.html
#scientificcomputing
#Sustainability
1/n
A really good summary on sustainability and ICT by Koen De Bosschere and Patrick Blouet. They discuss the basics of climate change, carbon emissions, embodied carbon, and discuss this for ICT: ”Sustainability is the single most important societal challenge of the 21st century, and
computing turns out to be both a part of the problem and a part of the solution.”:
https://zenodo.org/records/10875127
2/n
What do we need to do? Electrify everything and produce all electricity in a renewable way. This reduces global energy demand by 20%. After an increase in energy efficiency of about 20%, this leaves us with needing to multiply electricity production by a small factor (Sweden about two, Germany less than four, USA much larger). Doing stuff like LLMs, or getting less energy efficiency gains delays the transition and thus causes suffering. Doing the opposite like cycling instead of using an electric car reduces suffering. As long as we have still dirty parts of the grid, we need to keep saving energy.
3/n
There’s a rebound effect for data centers (Masanet 20). Overall, the rebound effect is hard to deal with. This is where changing the paradigm from ”time to solution” to ”energy to solution” could be helpful. Additionally, energy labeling of applications would be great, like for food (great talk by Kerstin Eder). Also, we need to train practitioners. Sustainable Carpentry as a branch of software carpentry? And include things in courses, starting with teaching basic units of energy and how much energy compute uses.
4/n
It is important to remind oneself that ”time to solution” is a good measure under the assumption that this is useful for science, so that if a job suddenly only takes half the time, then we can do double the science. Otherwise put, a lot of gains can already be obtained on the modelling side, by doing it ”smarter”.
5/n
If one then says one wants to look at ”energy to solution” instead, then this is equivalent to the other measure if we have perfect load balancing and intense usage of all cores. If not, then there are differences and something can be done.
6/n
