How accurate is the SIMPLE Law model?

It achieved R2=0.969, an MAPE of 11.97%, and 3.1 to 3.5× lower error than prior state-of-the-art, with no per-model calibration.

Does prefix caching eliminate the agentic energy overhead?

No. Prefix caching cuts prefill energy by 93% but leaves the KV-cache term unaffected, so cached agentic requests still cost 4.8× more than non-agentic ones.

What happens to Intelligence per Joule as reasoning budget increases?

Intelligence per Joule peaks at an intermediate reasoning budget and declines by up to 3.14× as that budget is scaled 4×.

Large Language ModelsAI Safety & AlignmentAI Weekly4h agoNov 6, 2026, 16:00 JST

SIMPLE Law finds 13.0× energy overhead for agentic LLM serving

Send on LINE

3 Key Points

What happened
University of Bologna and CINECA researchers published The SIMPLE Law for Energy Cost of LLM Inference, a four-parameter model validated on 13 dense LLMs from six model families on NVIDIA H100 GPUs with no per-model calibration, reaching R2=0.969 and finding a 13.0× energy overhead for agentic serving.
Why it matters
The model lets infrastructure teams estimate inference energy from workload parameters instead of running per-model benchmarks, and prefix caching cuts prefill energy 93% but leaves the KV-cache term unaffected.

WHO IT HITSInfrastructure and platform teams that run multi-turn and tool-calling AI workloads can now estimate energy from workload parameters rather than per-model benchmarking, which may change how they plan capacity and cost.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

Context & Analysis

The SIMPLE Law separates prefill energy from decode energy, and splits decode into KV-cache cost and next-token computation. That decomposition is what allows the model to be validated across 13 dense LLMs from six model families on NVIDIA H100 GPUs without per-model calibration, reaching R2=0.969 and an MAPE of 11.97%. The paper was posted to arXiv on November 2, 2026 by Hiari Pizzini Cavagna, Andrea Proia, Daniele Cesarini, Zeynep Kiziltan, and Andrea Bartolini.

The finding on reasoning budgets adds a second dimension: Intelligence per Joule peaks at an intermediate budget and declines by up to 3.14× as that budget is scaled 4×. That suggests extending reasoning past a moderate level costs progressively more energy for smaller accuracy gains, a trade-off infrastructure teams may need to weigh when configuring serving workloads.

FAQ

How accurate is the SIMPLE Law model?: It achieved R2=0.969, an MAPE of 11.97%, and 3.1 to 3.5× lower error than prior state-of-the-art, with no per-model calibration.
Does prefix caching eliminate the agentic energy overhead?: No. Prefix caching cuts prefill energy by 93% but leaves the KV-cache term unaffected, so cached agentic requests still cost 4.8× more than non-agentic ones.
What happens to Intelligence per Joule as reasoning budget increases?: Intelligence per Joule peaks at an intermediate reasoning budget and declines by up to 3.14× as that budget is scaled 4×.

AI WeeklyRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.