Profiling is only as useful as it is accurate.
For example, when an optimizer evaluates candidate configurations, biased or noisy measurements can steer it toward the wrong answer.
In this post, we describe the GPU energy profiling methodology adopted in Kareus, show experimentally why it matters, and introduce the zeus.profile module that automates the process.
The ML.ENERGY Leaderboard went from v2.0 (September 2024) to v3.0 (December 2025) with major changes: up-to-date models, hardware, software, and datasets. The v3.0 blog post covered the details of the v3.0 results, but how to they compare to the times v2.0? Are we making progress on energy efficiency? In this short post, we would like to look at the impact of software optimizations on energy efficiency over time, using the Llama 3.1 family as a case study.
With The ML.ENERGY Benchmark v3.0 we released in December 2025, we expanded our scope to up-to-date important models, tasks, and GPU hardware. This included 46 models across 7 tasks, producing 1,858 configurations on NVIDIA H100 and B200 GPUs.1 As always, latest benchmarking results are public and can be browsed on The ML.ENERGY Leaderboard.
In this post, we first present empirical observations from measurements, and then develop a reasoning framework that explains why we observe certain energy behaviors.
If you want to see how much energy LLM inference consumes on Apple Silicon, it's hard to find a straightforward way to do this programmatically, from within code. In this post, we'll explore how we can do this.
Zeus's batch size optimizer changes the model's training batch size to optimize time and energy consumption, but does that hurt the model's final quality? Short answer: No. Let's look into how in today's post.
To optimize something, you need to be able to measure it right. In this post, we'll look into potential pitfalls and best practices for GPU energy measurement.