https://carbonforge.ai/blog Field notes on AI inference at the operating point. Measurement, methodology, and observations on tokens per watt in production. en Wed, 03 Jun 2026 13:48:29 GMT 2026-06-03T13:48:29Z en https://carbonforge.ai/blog/adaptive-sm-clocking-for-energy-efficient-llm-serving <div class="hs-featured-image-wrapper"> <a href="https://carbonforge.ai/blog/adaptive-sm-clocking-for-energy-efficient-llm-serving" title="" class="hs-featured-image-link"> <img src="https://carbonforge.ai/hubfs/blog/Memory-bound%20decode-less%20power-no%20throughput%20loss.png" alt="Adaptive SM clocking for energy-efficient LLM serving" class="hs-featured-image" style="width:auto !important; max-width:50%; float:left; margin:0 15px 15px 0;"> </a> </div> <p>LLM serving is moving from occasional jobs to persistent service infrastructure. At that scale, GPU power is no longer a background detail. It sets thermal limits, caps how much inference fits on a given site, and decides where new capacity can be deployed <a href="https://powering-intelligence.epri.com/executive-summary.html">[15]</a>. Integrated over time, that draw becomes energy, and energy is what drives operating cost <a href="https://buildings.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report">[14]</a>.</p> <div class="hs-featured-image-wrapper"> <a href="https://carbonforge.ai/blog/adaptive-sm-clocking-for-energy-efficient-llm-serving" title="" class="hs-featured-image-link"> <img src="https://carbonforge.ai/hubfs/blog/Memory-bound%20decode-less%20power-no%20throughput%20loss.png" alt="Adaptive SM clocking for energy-efficient LLM serving" class="hs-featured-image" style="width:auto !important; max-width:50%; float:left; margin:0 15px 15px 0;"> </a> </div> <p>LLM serving is moving from occasional jobs to persistent service infrastructure. At that scale, GPU power is no longer a background detail. It sets thermal limits, caps how much inference fits on a given site, and decides where new capacity can be deployed <a href="https://powering-intelligence.epri.com/executive-summary.html">[15]</a>. Integrated over time, that draw becomes energy, and energy is what drives operating cost <a href="https://buildings.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report">[14]</a>.</p> <img src="https://track-na3.hubspot.com/__ptq.gif?a=342884351&amp;k=14&amp;r=https%3A%2F%2Fcarbonforge.ai%2Fblog%2Fadaptive-sm-clocking-for-energy-efficient-llm-serving&amp;bu=https%253A%252F%252Fcarbonforge.ai%252Fblog&amp;bvt=rss" alt="" width="1" height="1" style="min-height:1px!important;width:1px!important;border-width:0!important;margin-top:0!important;margin-bottom:0!important;margin-right:0!important;margin-left:0!important;padding-top:0!important;padding-bottom:0!important;padding-right:0!important;padding-left:0!important; "> Tokens per watt LLM inference GPU power Mon, 01 Jun 2026 17:18:12 GMT jml@carbonforge.ai (Jean-Maxime Larouche) https://carbonforge.ai/blog/adaptive-sm-clocking-for-energy-efficient-llm-serving 2026-06-01T17:18:12Z