Green Code or Costly Code? Evaluating the Energy Costs of Iterative LLM-Based Code Optimizations

Sadra Emami; Benjamin Weigell; Mohammad Reza Mousavi; Gunel Jahangirova; Hugo Araujo; Bernhard Bauer

cs.SEOct 23, 2026

Green Code or Costly Code? Evaluating the Energy Costs of Iterative LLM-Based Code Optimizations

Authors: Sadra Emami, Benjamin Weigell, Mohammad Reza Mousavi, Gunel Jahangirova, Hugo Araujo, Bernhard Bauer

Organizations: University of Augsburg, Augsburg, Germany · King’s College London, London, United Kingdom

Abstract

Large Language Models (LLMs) are increasingly used to optimize code for performance. This practice often assumes that improvements in execution speed naturally translate into overall lower energy consumption. We challenge this assumption empirically by measuring whether performance-optimized code produced by LLM self-refinement pipelines is also energy-efficient relative to canonical human reference solutions, and whether the energy cost of the optimization process can be offset by subsequent execution-time savings. We evaluate 11 open-source models across four decoding configurations on all 1,000 problems of EffiBench, a state-of-the-art performance benchmark, while collecting per-process, hardware-level energy measurements throughout. Inference energy spans 0.79 to 8.68 MJ across model-variant pairs, which is roughly equivalent to the energy of 250 to 2800 GPT-4.1 queries. Limiting the maximum number of output tokens has the largest impact, reducing energy use by up to 69% compared to the default configuration. On the execution side, 42 of 44 configurations produce code that consumes more energy per problem than the canonical human reference, with overheads ranging from +4% to +886%. The two exceptions achieve only marginal gains, with break-even points on the order of 10^6 to 10^7 executions. Mixture-of-Experts architectures consume substantially less inference energy than equivalently-sized dense models without sacrificing correctness, making architectural choice a more consequential sustainability lever than decoding parameters. Performance optimization and energy efficiency are not always aligned goals: LLM self-refinement consistently produces faster-seeking but energy-heavier code relative to what a human engineer would write. All artifacts are available at https://doi.org/10.5281/zenodo.23598170.

Figures & tables

Group	Task	Inference Energy	Runtime Energy	Standardized Benchmark	Parameter Sweep	Break-Even Analysis	Iterative Pipeline	Canonical Reference
Inference-energy studies [ 12 , 5 , 13 , 14 ]	General LLM	●	○	○	◐	○	○	○
DL framework energy [ 15 ]	General DL	●	○	●	○	○	○	○
Inference energy for code gen. [ 16 , 17 , 18 , 19 ]	Code Generation	●	○	◐	○	○	○	○
Generated-code energy [ 10 , 7 , 20 , 21 ]	Code Generation	○	●	●	○	○	○	●
Code-optimisation pipelines [ 3 , 22 , 23 ]	Code Optimization	○	●	◐	○	○	●	◐
Break-even analysis [ 9 ]	Code Optimization	●	●	●	○	●	◐	○

TABLE I: Coverage of the dimensions relevant to our three RQs across prior work, grouped by capability profile. ○ = not addressed, ◐ = partial, ● = addressed.

Family	Model	Parameters	Architecture
Qwen3.5	Qwen3.5-4B	4B	Dense
	Qwen3.5-9B	9B	Dense
	Qwen3.5-27B	27B	Dense
	Qwen3.5-35B-A3B	35B	MoE
GLM	GLM-4.7-Flash	31B	Dense
GLM	GLM-4.6V-Flash	10B	Dense

TABLE II: Models evaluated in this work.

Model	ET (s)	NET	MU (MB)	NMU	TMU	NTMU	pass@1 (%)
GLM-4.6V-Flash	0.133	1.052	23.098	1.003	2.215	1.053	25.4
GLM-4.6V-Flash	0.131 (1.4%)	1.038 (1.4%)	23.091 (0.0%)	1.003 (0.0%)	2.189 (1.2%)	1.040 (1.3%)	25.4
GLM-4.7-Flash	0.129	1.034	23.150	1.005	2.168	1.039	25.0
GLM-4.7-Flash	0.123 (4.7%)	0.984 (4.8%)	23.133 (0.1%)	1.005 (0.1%)	2.062 (4.9%)	0.989 (4.8%)	25.0
Ministral-3B	0.126	0.994	23.439	1.018	2.170	1.024	32.7
Ministral-3B	0.120 (4.8%)	0.946 (4.9%)	23.248 (0.8%)	1.009 (0.8%)	2.026 (6.7%)	0.956 (6.7%)	32.7

TABLE III: Resource efficiency before (top row) and after (bottom row) self-optimization and pass@1, with relative change in parentheses. ET: Execution Time (s); MU: Max Memory Usage (MB); TMU: Total Memory Usage (MB

\cdot

s). NET, NMU and NTMU are ET, MU and TMU normalized by the canonical solution; values below 1.0 beat the canonical reference.

Model	$E_{\mathrm{total}}$	$E_{\mathrm{CPU}}$	$E_{\mathrm{GPU}}$	CPU %	GPU %
GLM-4.6V-Flash	1.926	1.231	0.695	63.9	36.1
GLM-4.7-Flash	6.077	3.636	2.440	59.8	40.2
Ministral-3B	1.273	0.829	0.444	65.1	34.8
Ministral-8B	2.625	1.693	0.932	64.5	35.5
Ministral-14B	3.512	2.220	1.291	63.2	36.8
Nemotron-Cascade	3.085	1.913	1.171	62.0	38.0

TABLE IV: Inference energy consumption per model during self-optimization, in megajoules (MJ). CPU % and GPU % are each component’s share of

E_{\mathrm{total}}

Fig. 1: Total inference energy

E_{\mathrm{total}}

(MJ) per model and decoding variant after five epochs of self-optimization. mt1024 is almost always the lowest-energy variant and t0.5 is highest in most cases.

Fig. 2: Correctness efficiency (pass@1 /

E_{\mathrm{total}}

) vs. NET for all model-variant pairs. Shaded region: optimal zone. Dashed line: canonical baseline (NET = 1.0). Nemotron-Nano mt1024 is clipped due to pass@1 collapse.

Model	Variant	NET	NMU	NTMU	pass@1 (%)	$E_{\mathrm{total}}$	CPU %	GPU %
GLM-4.6V-Flash	default	1.038	1.003	1.040	25.4	1.930	63.9	36.1
	ctx65536	1.031	1.002	1.033	26.1	1.790	61.5	38.5
	mt1024	1.059	1.002	1.058	9.0	0.790	61.2	38.8
	t0.5	1.028	1.003	1.028	25.4	1.850	61.7	38.3
GLM-4.7-Flash	default	0.984	1.005	0.989	25.0	6.080	59.8	40.2
	ctx65536	0.966	1.004	0.971	25.5	6.110	60.2	39.8

TABLE V: Resource efficiency, inference energy, and pass@1 after self-optimization. N-prefixed metrics are normalized against the canonical solution; values below 1.0 indicate better performance than the canonical reference.

E_{\mathrm{total}}

in MJ.

Fig. 3: Left:

(E_{\mathrm{canon}}-E_{\mathrm{opt}})/n

; points left of zero beat the canonical baseline. Right:

(E_{\mathrm{base}}-E_{\mathrm{opt}})/n

; positive values indicate improvement over the base epoch.

Model	$E_{\mathrm{total}}$ (MJ)	gCO 2 eq
GLM-4.6V-Flash	1.926	116.1
GLM-4.7-Flash	6.077	366.3
Ministral-3B	1.273	76.8
Ministral-8B	2.625	158.2
Ministral-14B	3.512	211.7
Nemotron-Cascade	3.085	186.0

TABLE VI: Carbon footprint of the five-epoch optimization loop (default variant), computed using

\gamma=217

gCO 2 eq/kWh [ 35 ] .

Large Language Model-Guided Optimization

Large Language Model-Guided Optimization uses language models to propose, steer, and refine solutions to hard search and design problems.

See all 600 papers

Oct 9, 2026

Measuring the Inference Energy Efficiency of Open-Weight LLMs across Benchmarks, Reasoning Modes, and Hardware

Benchmark scores tell us what an LLM can do, but not how much energy is required to deliver that capability. We directly measure this trade-off for 86 open-weight model checkpoints, evaluated on seven benchmarks and six hardware platforms ranging from a consumer GPU to an eight-GPU server. We find large differences in energy efficiency even among models with similar capability. Most configurations lie away from the score--energy Pareto frontier: they use a median of 2.6--4.8

\times

more energy than another configuration that scores at least as well, with differences reaching 27

\times

. Reasoning can improve capability, but enabling thinking increases energy by about 3.7

\times

on average, and the benefit depends strongly on the model family and output budget. Generated-token count is also an imperfect proxy for energy, misordering 18--24% of model pairs. In contrast, FP8 reduces energy by roughly a quarter with little change in score, while changing hardware largely rescales energy without changing the relative ranking of models. These results show that capability alone gives an incomplete picture of efficiency. We argue that benchmark scores should be reported together with measured energy consumption under a specified serving regime and output budget.

Zeyu Yang, Wes Armour

Department of Engineering Science, University of Oxford

Energy-Aware LLM Inference LLM Inference Efficiency Energy-Efficient ML Open-Weight Language Models LLM Evaluation Efficient Language Model Reasoning

Ask PDF

Oct 1, 2026

Improving the Energy-Efficiency of the Code Generated by LLMs through Effective Prompting

As AI-assisted programming becomes increasingly mainstream, the environmental impact of AI-generated software has emerged as an important consideration. This motivates evaluating LLM-generated code beyond functional correctness by considering execution efficiency and energy consumption. However, despite substantial advances in code generation, frontier LLMs are rarely evaluated based on the energy efficiency of the code they produce. In this work, we conduct a comprehensive evaluation of 21 prompting strategies for energy-efficient code generation and identify 8 strategies for evaluation across 10 widely used open-weight and proprietary LLMs. We evaluate their effectiveness for both Python and C++ code generation relative to a baseline prompt. Across the evaluated models, the selected prompting strategies achieved energy reductions of up to 25% for Python and 17% for C++ code generation. At the model level, Python energy reductions reached up to 50% for Granite-4.0-H-Small, 39% for Claude 4.5 Haiku, and 28% for MiniMax M3, while C++ reductions reached up to 57% for Granite-4.0-H-Small and 6.4% for Qwen3-Coder-480B-A35B-Instruct. These results demonstrate that prompting strategies can substantially influence the energy consumption of LLM-generated code, although their effectiveness varies across models and programming languages. Our findings highlight the importance of incorporating energy efficiency into the evaluation and optimization of LLM-based code generation and provide practical insights into designing prompts for more sustainable AI-assisted programming.

Ritika Rekhi, Bing Zhang, Md Arman Islam +7

University at Buffalo, Buffalo, NY, USA · National Center for Supercomputing Applications, Urbana, IL, USA · IBM Research, Yorktown Heights, NY, USA

LLM Code Generation Code Generation LLM Prompting Code Generation Evaluation

Ask PDF

Apr 29, 2026

LLM-Guided Runtime Parameter Optimization for Energy-Efficient Model Inference

Large Language Models (LLMs) have become an integral part of many real-world workflows. However, LLMs consume a lot of energy, which becomes a large concern in the scale of the demand for these tools. As LLMs become integrated into different workflows, different applications have arisen to deal with the challenge of running inference for these tools. This raises another issue of choosing the runtime parameter values for these services in order to minimize the energy consumption. Oftentimes this requires deep knowledge of the application or traditional optimization methods that can take days to find optimal values. In this work, we created a human-in-the-loop flow with LLM-assisted runtime parameter optimization in order to solve this issue. With human-created, specific feedback prompting methods, chat-based LLMs can iteratively find energy-efficient inference parameters faster than traditional search methods. LLMs can also tailor their solutions to different hardware setups and easily take into account other system constraints. The enhanced prompt template was able to converge below the threshold at an average of 3.4 prompts compared to the baseline, which converged in an average of 5.2 prompts, and consistently achieved lower final energy per token. The enhanced prompt template also outperformed Sobol sampling in convergence speed.

Katelyn Crumpacker, Dimitrios Nikolopoulos

Virginia Polytechnic and State University

LLM Inference Efficiency Energy-Efficient ML Inference-Time Optimization Large Language Model-Guided Optimization Energy-Aware LLM Inference Power-Efficient Inference

Ask PDF

Abstract

Large Language Model-Guided Optimization

Explore similar work

Measuring the Inference Energy Efficiency of Open-Weight LLMs across Benchmarks, Reasoning Modes, and Hardware

Improving the Energy-Efficiency of the Code Generated by LLMs through Effective Prompting

LLM-Guided Runtime Parameter Optimization for Energy-Efficient Model Inference