Wim🧮<p>"Minimal Computing" by the Digital Humanities Climate Coalition</p><p><a href="https://sas-dhrh.github.io/dhcc-toolkit/toolkit/minimal-computing.html" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">sas-dhrh.github.io/dhcc-toolki</span><span class="invisible">t/toolkit/minimal-computing.html</span></a></p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
eupolicy.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
This Mastodon server is a friendly and respectful discussion space for people working in areas related to EU policy. When you request to create an account, please tell us something about you.
Administered by:
Server stats:
207active users
eupolicy.social: About · Profiles directory · Privacy policy
Mastodon: About · Get the app · Keyboard shortcuts · View source code · v4.4.3
#frugalcomputing
0 posts · 0 participants · 0 posts today
Stephen<p>I love <a href="https://mastodon.social/tags/Remarkjs" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Remarkjs</span></a> for creating <a href="https://mastodon.social/tags/slides" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>slides</span></a> from <a href="https://mastodon.social/tags/Markdown" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Markdown</span></a>, that is, until I noticed the library is ~650 kB. So, then I went full on <a href="https://mastodon.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a> and looked for the smallest <a href="https://mastodon.social/tags/HTML" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>HTML</span></a> <a href="https://mastodon.social/tags/presentation" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>presentation</span></a> library...</p><p>That honour, AFAICT, goes to Mark Dalgleish's modular <a href="https://mastodon.social/tags/Bespokejs" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Bespokejs</span></a>. The <a href="https://mastodon.social/tags/Yeoman" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Yeoman</span></a> generator failed for me, but with a little perseverance I hand coded a demo:<br><a href="https://blog.harlow.net.nz/presentations/" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">blog.harlow.net.nz/presentatio</span><span class="invisible">ns/</span></a><br>that only required 8 kB of <a href="https://mastodon.social/tags/Javascript" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>Javascript</span></a>. Remarkably, that's 1/80th of Remark.js!</p><p>👉 <a href="https://markdalgleish.com/projects/bespoke.js/" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">markdalgleish.com/projects/bes</span><span class="invisible">poke.js/</span></a></p>
DamonHD<p>Switched some load-profile graphs I produce from <a href="https://mastodon.social/tags/PNG" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>PNG</span></a> to <a href="https://mastodon.social/tags/SVG" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>SVG</span></a> for about an order of magnitude fewer bytes on the wire ... Almost small enough to inline... <a href="https://mastodon.social/tags/frugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>frugalComputing</span></a> </p><p><a href="https://www.earth.org.uk/note-on-site-technicals-100.html#2025-09-10" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://www.</span><span class="ellipsis">earth.org.uk/note-on-site-tech</span><span class="invisible">nicals-100.html#2025-09-10</span></a></p>
Wim🧮<p>To illustrate this, here is a graph from the paper. It shows the number of years of growth you can compensate through saving of emissions by any means except reducing the growth.<br>If your savings are <10%, you can compensate a few months. If it's around 50%, you can compensate a few years of growth. To compensate order of 10 years of growth you need to reduce emissions by more than 80%.<br>(4/4)</p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>When you take all those factors into account, it turns out the gain is so small that investing effort into this is purely a distraction from the bigger problem of the growth in data centre capacity.</p><p>(The paper I'm working on now tries to quantifies what that growth means in terms of global emissions. )<br>(3/4)<br><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>The idea of is that moving work from a a region with fossil fuel generation to one with renewables will reduce emissions. <br>This is of course correct. But it depends a lot on <br>- how much work you can move<br>- how much of the time you can move it<br>- what the difference in carbon intensity is between the high-CO2 and low-CO2 region<br>- what the contribution of the embodied carbon of the data centre is.<br>(2/4)<br><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>I wrote a paper that I am proud of: </p><p>"Modelling Scenarios for Carbon-aware Geographic Load Shifting of Compute Workloads"</p><p>It evaluates the scope for emission reductions of moving work between data centres in high-emission regions and low-emission regions.</p><p>tl;dr: the potential emission savings are no more than a few percent.</p><p>I have submitted it to a journal, it's under review but you can read the preprint here: <a href="https://arxiv.org/abs/2509.07043" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="">arxiv.org/abs/2509.07043</span><span class="invisible"></span></a></p><p>(1/4)<br><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Jonathan Schofield<p><span class="h-card" translate="no"><a href="https://mastodon.social/@grimalkina" class="u-url mention" rel="nofollow noopener" target="_blank">@<span>grimalkina</span></a></span> perhaps <span class="h-card" translate="no"><a href="https://scholar.social/@wim_v12e" class="u-url mention" rel="nofollow noopener" target="_blank">@<span>wim_v12e</span></a></span> is in the right ballpark? </p><p>More generally, maybe check out <a href="https://mastodon.social/tags/frugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>frugalComputing</span></a> </p><p><a href="https://mastodon.social/@urlyman/114223143337500275" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">mastodon.social/@urlyman/11422</span><span class="invisible">3143337500275</span></a></p>
Wim🧮<p>The paper by @maswan and myself</p><p> "Life Cycle Analysis for Emissions of Scientific Computing Centres" </p><p>has been published!</p><p>We develop a detailed model for the LCA of (HPC) data centres, including embodied carbon, server replacement and expansion. It is also applicable to other data centres. We also share the source code.</p><p>It shows how important embodied carbon becomes when the grid has more renewables.</p><p><a href="https://link.springer.com/article/10.1140/epjc/s10052-025-14650-8" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">link.springer.com/article/10.1</span><span class="invisible">140/epjc/s10052-025-14650-8</span></a></p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>At the risk of restating the obvious: the AI companies are spending $100 billion in a single quarter on data centres because they plan to use them at capacity to maximise profits. Therefore the actual energy and water cost of an individual query is irrelevant: the purpose of installing >10 GW of new data centre capacity per year globally is to use it fully. Ditto for the generation capacity needed to power them: it will be used, if not for data centres, then for something else.</p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>The purpose of Google's paper on energy and water consumption of their LLMs is to absolve individual users of their responsibility for the overall emissions and water usage. That way, they hope adoption will rise or at least not drop. The need the adoption to justify continued huge investments in data centres. And the damage of this investment is done even if the AI bubble would burst tomorrow. </p><p><a href="https://limited.systems/articles/the-real-problem-with-AI/" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">limited.systems/articles/the-r</span><span class="invisible">eal-problem-with-AI/</span></a><br><a href="https://limited.systems/articles/the-insatiable-hunger-of-openai/" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">limited.systems/articles/the-i</span><span class="invisible">nsatiable-hunger-of-openai/</span></a></p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>My paper with @maswan </p><p> "Life Cycle Analysis for Emissions of Scientific Computing Centres" </p><p>has been accepted fro publication in European Physical Journal C !</p><p>We develop a detailed model for the LCA of (HPC) data centres, including embodied carbon, server replacement and expansion. It is also applicable to other data centres. We also share the source code.</p><p>It shows how important embodied carbon becomes when the grid has more renewables.</p><p><a href="https://arxiv.org/abs/2506.14365" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="">arxiv.org/abs/2506.14365</span><span class="invisible"></span></a></p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>I'm doing a deep dive into the embodied carbon of GPU servers for AI training and inference, such as the nvidia DGX-A100 and DGX-B300. I was surprised to find that the embodied carbon is dominated by the RAM (i.e it's more than 50% of the total). </p><p>And the big difference between the DGX-A100 (2020) and the DGX-B300 (2025) is the amount of RAM. The embodied carbon of that RAM has more than doubled. </p><p>Second largest contribution is the SSD, but that has not increased in size.</p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p><span class="h-card" translate="no"><a href="https://mastodon.online/@frebelt" class="u-url mention" rel="nofollow noopener" target="_blank">@<span>frebelt</span></a></span> </p><p>My own estimate (<a href="https://wimvanderbauwhede.codeberg.page/articles/google-search-vs-chatgpt-emissions/" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">wimvanderbauwhede.codeberg.pag</span><span class="invisible">e/articles/google-search-vs-chatgpt-emissions/</span></a>) is between 0.002 kWh and 0.005 kWh per query. But that is for short queries (<100 words), and also this is for GPT-3.5. For GPT-4, it is likely to be 3x higher.</p><p>So all in all I would say the estimate for water could be up to 10x larger than the figure in the study, so it could vary between 10 and 500 ml per query with a response of order of 100 words. I'd say a good ballpark figure would therefore be 50 ml per query.</p><p>(2/2) <a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>In short, not only are more chips made, the embodied carbon per chip is increasing as well.<br>(3/3)<br><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>Unfortunately, this is not what happens. Instead, the chips scale in performance and in storage density as the industry wants to keep Moore's law alive. So the the increased density does not lead to reduced embodied emissions. In other words, the embodied emissions per die have increased by 3x in that period, despite the much higher densities. </p><p>This is further compounded by the growth in wafer volume. Because of the AI hype, this growth is currently truly exponential.<br>(2/3)<br><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>I looked into the growth in embodied carbon (emissions from manufacturing) of the chips, those CPUs, GPUs, RAM and SSDs that fuel the AI hype. </p><p>For the same silicon area, the embodied carbon of chip production has grown 4x since 2010. At the same time, the density has increased about 100x.</p><p>At first sight, this is wonderful: if the compute performance and storage capacity would have remained the same as in 2010, the embodied carbon of every chip sold would be 25x lower. <br>(1/3)<br><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a></p>
Wim🧮<p>So we are looking at a growth of close to 20% which follows the medium-range scenario set out by McKinsey as discussed in my post:<br> <br><a href="https://wimvanderbauwhede.codeberg.page/articles/the-real-problem-with-AI/" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">wimvanderbauwhede.codeberg.pag</span><span class="invisible">e/articles/the-real-problem-with-AI/</span></a></p><p>(2/2)</p><p><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a> <br> <a href="https://scholar.social/tags/AIHype" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>AIHype</span></a> <a href="https://scholar.social/tags/ClimateEmergency" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>ClimateEmergency</span></a></p>
Wim🧮<p>This post shows a graph of capex for meta, google, microsoft and amazon<br><a href="https://mastodon.social/@samim/114957102999706720" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">mastodon.social/@samim/1149571</span><span class="invisible">02999706720</span></a></p><p>In the last quarter, it was >$100 billion. I checked those numbers, and the order of magnitude is certainly correct.</p><p>If all that goes to AI data centres, with a capex of $5 billion for a 500 MW data centre (from <a href="https://ravenscraig.co.uk/news/news-post/ravenscraig-plans-to-transform-former-steelworks-into-one-of-the-uks-largest-green-ai-data-centres/" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://</span><span class="ellipsis">ravenscraig.co.uk/news/news-po</span><span class="invisible">st/ravenscraig-plans-to-transform-former-steelworks-into-one-of-the-uks-largest-green-ai-data-centres/</span></a>) that means 10 GW of data centre capacity. Over a year it will be at least double so 20 GW<br>In 2023, global capacity was 55 GW. <br>(1/2)<br><a href="https://scholar.social/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a><br><a href="https://scholar.social/tags/AIHype" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>AIHype</span></a></p>
Earth Notes<p>RSS Podcast Feed Inefficiency - Climate cost of handling feeds ineptly... <a href="https://mastodon.energy/tags/RSS" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>RSS</span></a> <a href="https://mastodon.energy/tags/podcast" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>podcast</span></a> <a href="https://mastodon.energy/tags/FrugalComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>FrugalComputing</span></a> <a href="https://mastodon.energy/tags/LowCarbonComputing" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>LowCarbonComputing</span></a> <a href="https://mastodon.energy/tags/greenSoftware" class="mention hashtag" rel="nofollow noopener" target="_blank">#<span>greenSoftware</span></a> - <a href="https://www.earth.org.uk/RSS-efficiency.html" rel="nofollow noopener" translate="no" target="_blank"><span class="invisible">https://www.</span><span class="ellipsis">earth.org.uk/RSS-efficiency.ht</span><span class="invisible">ml</span></a></p>
TrendingLive feeds
Mastodon is the best way to keep up with what's happening.
Follow anyone across the fediverse and see it all in chronological order. No algorithms, ads, or clickbait in sight.
Create accountLoginDrag & drop to upload