acceptodds
Under review as a conference paper at ICLR 2027

Beyond Energy: When Sustainability Dimensions Reshape LLM Serving Decisions

Abstract

Large language model (LLM) serving has environmental impacts across energy consumption, carbon emission, water consumption, and biodiversity loss. Yet these dimensions are largely evaluated in isolation, leaving it unclear when and how they lead to different optimization decisions. We present PRISM, a unified framework for characterizing and optimizing LLM serving across energy, carbon, water, and biodiversity impacts. Our analysis reveals a fundamental distinction: computing configurations determine energy consumption, whereas where and when LLM serving is deployed determine its carbon, water, and biodiversity impacts. Under a fixed deployment choice and operational-only accounting, all dimensions preserve the same energy-based configuration ranking. Deployment rankings can diverge across dimensions, while embodied impacts can break configuration invariance when they exceed a lifecycle crossover boundary. PRISM identifies these conditions, quantifies cross-dimensional regrets, and balances the four dimensions. In regional-routing experiments, PRISM reduces median worst-case regret by 50.2% relative to the strongest baseline.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Large Language Model Serving, Environmental Impact, Energy, Carbon, Water, Biodiversity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nWEWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

PRISM combines measured or modeled IT energy with regional environmental intensities and allocated hardware-manufacturing impacts to characterize energy, carbon, water, and biodiversity consequences of LLM serving. It distinguishes configuration selection from regional placement, derives ranking invariance under a separable operational model and a lifecycle crossover condition, and solves normalized minimax-regret routing problems. The evaluation includes offline regional routing, rolling-horizon routing, and a joint configuration-region extension with agent completion time. The main reported result is a 50.2% median reduction in worst-case regret relative to the adapted WaterWise baseline.

Critical strengths

The paper evaluates an important trade-off with explicit functional units

There is meaningful implementation and sensitivity work

Critical weaknesses

a) The claimed decision theory and optimization evidence do not yet establish a substantial methodological advance. Under Eq. 2 , operational impacts are positive scalar multiples of the same energy, making Propositions 1-2 consequences of the assumed model; the crossover condition is the corresponding rearrangement after adding emodied terms. Likewise, an exact four-dimensional minimax solver is expected to perform best on its own worst-case-regret objective against single-objective or carbon-water policies. The 50.2% result quantifies the effect of changing the decision objective, but does not by itself demonstrate a superior optimization method.

b) The water-accounting formulas are not equivalent as written, and the reported uncertainty test does not resolve this issue. Tab. 1 is repeated in Tab. 3, uses [WUE(r,t)+PUE(r,t) EWIF(r,t)] WSF(r), applying the datacenter-region stress factor to electricity-related water. Appx A.1 (i.e., Eq 14 -15) instead weights each electricity-related water flow by WSF at its location, explicitly not uniformly at the datacenter region; EWIF is defined there as the unweighted total water intensity. These formulations require an additional co-location/equal-stress assumption or a different EWIF definition to agree. The distinction can affect regional water rankings and the 4.637% regret headline. The rolling-horizon 20% noise experiment perturbs future inputs while observing the current hour exactly and knowing deman/capacity; it does not validate the underlying water-flow geography or current environmental coefficients (Appx. A.7.1).

Questions for authors

a) Which water formula is implemented? Please reconcile Tab. 1/3 with Eqs. 14 & 15, and identify the locations assigned to electricity-related water flow, and show whether the main water-regret and routing conclusions persist under the consistent formulation.

b) Under the same normalized minimax decision rule, what consequential decisions change when biodiversity is added to energy/carbon/water? A matched-objective dimensional ablation would isolate the benefit of the fourth dimension from the benefit of replacing a weighted-sum objective with minimax regret.

c) How stable are the selected regions and the headline regret improvement to uncertainty in the current water-stress/electricity-water/biodiversity coefficients, rather than only independent future forecast noise? A bounded sensitivity analysis on the existing routing instance would materially affect my assessment.

6 comments
AuthorsOpenReview ↗

Response to Reviewer nWEW (Part 1)

Thank you for the detailed comments. We have clarified the contribution, explained the water formula used in the implementation, and added the two requested experiments.

The new comparisons separate the effect of the decision rule from the effect of adding biodiversity. The sensitivity analysis tests uncertainty in current electricity-related coefficients, rather than only future forecast errors. We explain the scope and limits of each test below. All changes are highlighted in the revised PDF.

1. Contribution of the propositions and optimization results

The propositions follow from the assumed model, and the 50.2% result does not by itself demonstrate a superior optimization method.

We agree. PRISM does not introduce a new accounting model or a new optimization algorithm. The 50.2% result measures the effect of changing the decision objective. It does not show that PRISM has a better solver. We have revised the text to make these claims explicit.

Under fixed deployment conditions, each operational impact is IT energy multiplied by a fixed impact intensity (Eq. 2). It is how operational impacts are defined in established accounting methods in both iterature and industry practice:

  • Carbon. Operational carbon is energy consumption times the carbon intensity of the supplying grid. This is the model used by Strubell et al. [1], the BLOOM footprint study [2], LLMCarbon (Eqs. 9–10) [3], and ACT (Eq. 2) [4].
  • Industry practice. Google applies the same form to carbon and on-site water in its production measurements of Gemini serving [5].
  • Water. Operational water follows the same structure, EIT(WUE+PUE⋅EWIF)E_{\mathrm{IT}}(\mathrm{WUE} + \mathrm{PUE}\cdot\mathrm{EWIF}) [6, 7].
  • Biodiversity. Biodiversity applies ReCiPe 2016 characterization factors to the same electricity-related flows [8].

PRISM’s contribution is the combined analysis of serving decisions across energy, carbon, water, and biodiversity. It compares these dimensions under common service requirements and accounting boundaries, identifies when they select different configurations or regions, and measures the resulting penalties.

PRISM uses an established minimax method to select a compromise. The optimization experiments show the tradeoffs produced by the selected objective, rather than solver superiority. The new comparison below also separates this effect from the effect of including biodiversity.

We have revised the contribution statement in Section 1 to make this scope clearer.

References

  • [1] E. Strubell, A. Ganesh, and A. McCallum. Energy and policy considerations for deep learning in NLP. ACL, 2019.
  • [2] A. S. Luccioni, S. Viguier, and A.-L. Ligozat. Estimating the carbon footprint of BLOOM, a 176B parameter language model. JMLR, 2023.
  • [3] A. Faiz et al. LLMCarbon: Modeling the end-to-end carbon footprint of large language models. ICLR, 2024.
  • [4] U. Gupta et al. ACT: Designing sustainable computer systems with an architectural carbon modeling tool. ISCA, 2022.
  • [5] C. Elsworth et al. Measuring the environmental impact of delivering AI at Google scale. arXiv:2508.15734, 2025.
  • [6] Y. Jiang, R. Kanakagiri, R. Basu Roy, and D. Tiwari. ThirstyFLOPS: Water footprint modeling and analysis toward sustainable HPC systems. SC, 2025.
  • [7] S. Ren, B. Tomlinson, R. W. Black, and A. W. Torrance. Reconciling the contrasting narratives on the environmental impact of large language models. Scientific Reports, 2024.
  • [8] M. A. J. Huijbregts et al. ReCiPe 2016: A harmonized life cycle impact assessment method at midpoint and endpoint level. RIVM, 2016.
  • [9] Y. Jiang, R. Basu Roy, R. Kanakagiri, and D. Tiwari. WaterWise: Co-optimizing carbon- and water-footprint toward environmentally sustainable cloud computing. PPoPP, 2025.
AuthorsOpenReview ↗

Response to Reviewer nWEW (Part 2)

2. Water formula used in the implementation

Which water formula is implemented? Please reconcile Tables 1/3 with Eqs. 14–15, identify the locations assigned to electricity-related water flows, and show whether the main conclusions persist.

The implementation uses the formula in Tables 1/3, a regional approximation of Eqs. 14–15. It applies one regional water-stress factor, WSF(r)\mathrm{WSF}(r), to both direct cooling water and electricity-related water: IW,op=EIT[WUE(r,t)+PUE(r,t) EWIF(r,t)]WSF(r).I_{W,\mathrm{op}}=E_{\mathrm{IT}}\left[\mathrm{WUE}(r,t)+\mathrm{PUE}(r,t)\,\mathrm{EWIF}(r,t)\right]\mathrm{WSF}(r).

Electricity-related water comes from the datacenter region's own grid. EWIF(r,t)\mathrm{EWIF}(r,t) combines the region's hourly generation mix [1] with per-source water-consumption factors [2], without electricity imports or exports. In Eqs. 14–15, ℓ\ell indexes where each water flow occurs, such as a power plant's watershed. Setting WSF(ℓ)=WSF(r)\mathrm{WSF}(\ell)=\mathrm{WSF}(r) for all of region rr's electricity-related water reduces Eqs. 14–15 to the formula above. The reviewer is correct that the paper did not state this condition.

We use this approximation because plant-level water consumption linked to source watersheds is not available across regions. Prior work makes the same choice: WaterWise [8] and ThirstyFLOPS [9] also apply the datacenter region's factor to both direct and electricity-related water. The water-stress factors come from AWARE 2.0 [3].

All results, including the 4,637% energy-to-water regret and the routing results, use this formula. We agree that the rolling-horizon forecast experiment does not test water-use locations; our response (part 3) addresses uncertainty in the current coefficients.

Appendix A.1 now states the location-based scope, defines ℓ\ell and rr, states the approximation WSF(ℓ)=WSF(r)\mathrm{WSF}(\ell)=\mathrm{WSF}(r), and derives the Table 1 formula from Eqs. 14–15 (new Eq. 16).

3. Decisions that change when biodiversity is added

Under the same normalized minimax decision rule, what consequential decisions change when biodiversity is added to energy, carbon, and water?

Adding biodiversity has almost no effect in the original 12-region experiment, but changes routing in a selected four-region set where biodiversity matters.

We compare PRISM restricted to energy, carbon, and water, called PRISM-ECW, with PRISM using all four dimensions. Both use the same minimax rule, workload, service constraints, and regional capacities. Only the set of dimensions changes.

In Figure 7 (12 regions, 24-hour Azure trace), carbon and water have the largest regrets and thus determine the tradeoff. Adding biodiversity changes worst-case regret and biodiversity regret by less than 0.02 percentage points in every capacity setting. Thus, biodiversity does not explain the reported improvement over WaterWise in Figure 7.

To test a case in which biodiversity can determine the worst-case regret, we select four regions from Table 10 using their measured environmental data: Sydney, Paris, Tel Aviv, and Johannesburg. The same Azure workload, hourly environmental data for June 24, 2024, capacity-generation procedure, and 20 capacity seeds are unchanged. We selected this set to test a case where biodiversity can affect the decision. The table reports median regret across the 20 capacity seeds. Regret is the increase above the best feasible value of each dimension.

PolicyEnergyCarbonWaterBiodiversityWorst-case
PRISM-ECW7.2%7.2%6.5%11.3%11.3%
PRISM7.6%6.5%2.0%7.6%7.6%

Adding biodiversity also changes where requests are served. We present the mean share of GPU demand for each region across the 20 settings for both policy:

PolicySydneyParisTel AvivJohannesburg
PRISM-ECW34.6%35.2%19.7%10.4%
PRISM33.6%35.2%27.8%3.5%

These are mean shares of GPU demand across the 20 capacity settings.

PRISM-ECW uses Johannesburg partly because its low PUE reduces energy use. However, Johannesburg has the highest modeled biodiversity intensity among these four regions. PRISM moves most of this demand to Tel Aviv, and Johannesburg's share falls in 19 of the 20 settings. Biodiversity regret falls from 11.3% to 7.6% and water regret from 6.5% to 2.0%, while energy regret rises from 7.2% to 7.6%. Worst-case regret falls in all 20 settings.

This is a concrete decision change: PRISM moves demand away from the region whose grid harms biodiversity most, at a small energy cost. PRISM finds from the data whether biodiversity matters, so operators do not need to know this in advance.

Section 7 now reports both results in Table 2. Appendix A.7.6 gives the per-dimension regrets (Table 15) and the routing change (Table 16).

AuthorsOpenReview ↗

Response to Reviewer nWEW (Part 3)

4. Sensitivity to uncertainty in current environmental coefficients

How stable are the selected regions and the headline regret improvement to uncertainty in current coefficients, rather than only future forecast noise?

The findings hold under uncertainty in the current environmental coefficients. These coefficients are the paper's operational impact intensities, and their uncertainty comes from the per-source electricity factors behind them.

Uncertainty model

Each regional intensity is the generation-weighted sum of per-source factors, F(r,t)=∑uϕu(r,t) FuF(r,t)=\sum_u \phi_u(r,t)\,F_u, where ϕu(r,t)\phi_u(r,t) is the generation share of energy source uu and FuF_u is its factor. We now state this equation in Appendix A.1 (Eq. 20).

Following RHODES [7], these per-source factors have data-availability uncertainty: each is published as a range over a few estimates, too few to fit a distribution. For example, hydropower water consumption ranges from 5.4 to 68.1 L/kWh across three estimates [4]. We therefore bound each factor by its published range.

  • Varied: carbon, water, NOx, and SO2 factors of 8 of the 10 energy sources (32 factors), with water ranges from [4] and the others from [5]. Biodiversity is recomputed from these through ReCiPe [6].
  • Fixed: generation shares [1], water-stress factors [3], PUE, and WUE, which have no published ranges.

One factor vector applies to all regions and hours. For Figure 6, we compute the exact range of each cross-regret over the full set by linear programming. We also evaluate 2,048 extreme factor combinations (stress points).

Regional choices and cross-dimensional regret

For annual intensities (Figure 6, left), most cross-regrets move above and below their point estimates but stay positive. The energy-to-water regret ranges from 2,303% to 6,888%, compared with 4,637% at the point estimates. Only the carbon–biodiversity and water–biodiversity regrets can reach zero, at 1.3% and 2.3% of the stress points. Each dimension's selected region (Melbourne, Los Angeles, Kuala Lumpur, and Frankfurt) also has the smallest worst-case regret over the full range, so the regional selection is stable.

For hourly intensities (Figure 6, right), the disagreement persists over the full range in all 72 hours for pairs involving energy, and in 64 and 60 hours for carbon–water and water–biodiversity.

Routing performance with fixed decisions

For Figure 7, we fix each policy's allocation and evaluate it at the 2,048 stress points, recomputing the four per-dimension optima at each point. Robust PRISM bounds the regret at 256 separate training combinations and remains a linear program. Ideal routing re-solves PRISM with the true factors known, which gives a lower bound for any policy. The table reports worst-case regret as the median over the 20 capacity settings.

PolicyPoint estimatesMedian over stress pointsMaximum over stress points
WaterWise157.8%151.9%252.0%
PRISM75.1%95.8%146.1%
Robust PRISM85.2%104.2%130.7%
Ideal routing—75.5%128.6%

PRISM keeps both statistics well below WaterWise, in 19 and 18 of the 20 settings. Much of its remaining regret is inherent: ideal routing has a median of 75.5%, and PRISM exceeds it by a median of 25.3 percentage points, compared with 76.2 for WaterWise. Robust PRISM lowers the maximum from 146.1% to 130.7% in all 20 settings, close to the 128.6% lower bound. In exchange, its regret is higher at the point estimates and at the median stress point, so operators can choose by risk preference.

Appendix A.1 adds the intensity equation. New Appendix A.6 gives the uncertainty model and the Figure 6 range table (Table 13), and Figure 6 now shows the hourly ranges as shaded bands. Sections 6 and 7 summarize the results, with routing in Table 2. Appendix A.8.1 gives robust PRISM and ideal routing (Table 17).

Reviewer_hZ9JOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PRISM, a unified framework for analyzing the environmental impacts of large language model (LLM) serving, covering energy consumption, carbon emissions, water consumption, and biodiversity impacts. The authors develop a unified accounting framework for both operational and embodied impacts from hardware manufacturing. They examine the consistency of computing configuration rankings under fixed deployment choices, as well as differences in deployment rankings across regions and time periods and the conditions under which embodied impacts alter configuration rankings. For regional request scheduling, the paper employs a minimax relative regret objective under capacity constraints to balance the four environmental dimensions. Through experiments involving diverse models, GPU platforms, serving workloads, and regional environmental data, the authors present findings on configuration selection, regional preferences, and multiobjective routing.

Critical strengths
  1. The paper analyzes the consistency of configuration rankings across different environmental dimensions and derives explicit conditions under which embodied impacts can reverse these rankings, providing useful insights into when energy based configuration selection remains valid.

  2. The experimental evaluation covers diverse models, GPU platforms, workloads, and deployment regions, revealing substantial differences in the deployment choices preferred by different environmental objectives and quantifying the resulting tradeoffs.

Critical weaknesses
  1. The main result on operational configuration ranking consistency largely follows from the assumption that environmental impacts are proportional to IT energy under a fixed deployment choice. Furthermore, the minimax relative regret objective is based on a standard minimax scalarization technique rather than introducing a novel optimization algorithm.

  2. The baseline comparison with WaterWise [1] is not entirely fair, as the authors simultaneously change both the number of optimization objectives and the scalarization method. Without a minimax baseline considering only carbon and water impacts, it is difficult to determine whether the reported improvements result from including additional environmental dimensions or simply from changing the scalarization method.

  3. Regional routing evaluation assumes that hardware availability and execution characteristics are identical across all regions. Incorporating regional variations and network latency into the assessment would help further enhance its applicability in real-world scenarios.

[1]WaterWise: Co-optimizing Carbon- and Water-Footprint Toward Environmentally Sustainable Cloud Computing.

Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer hZ9J (Part 1)

Thank you for these comments. They helped us clarify the contribution and improve the evaluation. We added experiments that separate the effect of the optimization method from the effect of including more environmental objectives. We also clarified which regional differences the evaluation includes and which remain outside its scope. All changes are highlighted in the revised PDF.

1. Contribution of the ranking result and the optimization method

The main result on operational configuration ranking consistency largely follows from the assumption that environmental impacts are proportional to IT energy under a fixed deployment choice. Furthermore, the minimax relative regret objective is based on a standard minimax scalarization technique rather than introducing a novel optimization algorithm.

To the best of our knowledge, PRISM is the first to show when energy is sufficient to guide serving decisions, when other environmental measures change those decisions, and how much these differences matter.

For a fixed region and time, the operational impact in each dimension equals IT energy multiplied by a fixed impact intensity (Eq. 2). It is how operational impacts are defined in established accounting methods in both iterature and industry practice:

  • Carbon. Operational carbon is energy consumption times the carbon intensity of the supplying grid. This is the model used by Strubell et al. [1], the BLOOM footprint study [2], LLMCarbon (Eqs. 9–10) [3], and ACT (Eq. 2) [4].
  • Industry practice. Google applies the same form to carbon and on-site water in its production measurements of Gemini serving [5].
  • Water. Operational water follows the same structure, EIT(WUE+PUE⋅EWIF)E_{\mathrm{IT}}(\mathrm{WUE} + \mathrm{PUE}\cdot\mathrm{EWIF}) [6, 7].
  • Biodiversity. Biodiversity applies ReCiPe 2016 characterization factors to the same electricity-related flows [8].

PRISM does not claim to introduce a new optimization algorithm. It uses minimax relative regret to balance the selected objectives. Instead, PRSIM is a new framework that connects common workload measurements, various environmental accounting, comparisons of decisions, and deployment selection.

We have revised the contribution statement in Section 1 to make this scope clearer.

References

  • [1] E. Strubell, A. Ganesh, and A. McCallum. Energy and policy considerations for deep learning in NLP. ACL, 2019.
  • [2] A. S. Luccioni, S. Viguier, and A.-L. Ligozat. Estimating the carbon footprint of BLOOM, a 176B parameter language model. JMLR, 2023.
  • [3] A. Faiz et al. LLMCarbon: Modeling the end-to-end carbon footprint of large language models. ICLR, 2024.
  • [4] U. Gupta et al. ACT: Designing sustainable computer systems with an architectural carbon modeling tool. ISCA, 2022.
  • [5] C. Elsworth et al. Measuring the environmental impact of delivering AI at Google scale. arXiv:2508.15734, 2025.
  • [6] Y. Jiang, R. Kanakagiri, R. Basu Roy, and D. Tiwari. ThirstyFLOPS: Water footprint modeling and analysis toward sustainable HPC systems. SC, 2025.
  • [7] S. Ren, B. Tomlinson, R. W. Black, and A. W. Torrance. Reconciling the contrasting narratives on the environmental impact of large language models. Scientific Reports, 2024.
  • [8] M. A. J. Huijbregts et al. ReCiPe 2016: A harmonized life cycle impact assessment method at midpoint and endpoint level. RIVM, 2016.
  • [9] Y. Jiang, R. Basu Roy, R. Kanakagiri, and D. Tiwari. WaterWise: Co-optimizing carbon- and water-footprint toward environmentally sustainable cloud computing. PPoPP, 2025.
AuthorsOpenReview ↗

Response to Reviewer hZ9J (Part 2)

2. Separating the optimization method from the number of objectives

The baseline comparison with WaterWise is not entirely fair, as the authors simultaneously change both the number of optimization objectives and the scalarization method. Without a minimax baseline considering only carbon and water impacts, it is difficult to determine whether the reported improvements result from including additional environmental dimensions or simply from changing the scalarization method.

We agree that the original comparison did not separate the two effects. We have added a new ablation showing that the gain comes from PRISM's formulation, not from the number of dimensions. A second experiment shows when the extra environmental dimensions help.

WaterWise minimizes a weighted sum of normalized carbon and water. PRISM instead bounds each dimension's regret relative to its own optimum, without weights (Eqs. 7 and 9); choosing this minimax formulation is our contribution. A carbon–water minimax-regret policy is not a prior method. Instead, it is PRISM restricted to two dimensions.

We use the Figure 7 instance: 12 regions, the 24-hour Azure trace, and 20 capacity settings. We restrict PRISM to carbon and water (PRISM-CW) and to energy, carbon, and water (PRISM-ECW). All policies share the same constraints and the same per-dimension optima, and each is evaluated on all four dimensions. Regret is the percentage increase above the best feasible value of each dimension, and worst-case regret is the largest of the four. The table reports medians over the 20 capacity seeds.

PolicyEnergyCarbonWaterBiodiversityMedian worst-case regret
WaterWise5.1%22.7%157.8%2.7%157.8%
PRISM-CW5.5%75.1%75.1%27.9%75.1%
PRISM-ECW5.5%75.1%75.1%27.9%75.1%
PRISM5.5%75.1%75.1%27.9%75.1%

With the same two dimensions as WaterWise, PRISM-CW lowers worst-case regret in all 20 capacity seeds, from 157.8% to 75.1% at the median. It accepts higher carbon regret to cap the much larger water regret. In this instance, carbon and water have equal and largest regrets in every setting, while energy and biodiversity stay at least 32 percentage points below them. Adding these two dimensions therefore leaves every regret unchanged to within 0.03 percentage points. This matches Figure 6 (right), where carbon and water disagree strongly in every hour across the same 12 regions.

To test a case in which biodiversity can determine the worst-case regret, we select four regions from Table 10 using their measured environmental data: Sydney, Paris, Tel Aviv, and Johannesburg. The workload, capacity procedure, and capacity settings are unchanged.

PolicyEnergyCarbonWaterBiodiversityWorst-case
WaterWise8.7%0.8%2.2%1.2%8.7%
PRISM-CW8.6%1.3%1.3%1.6%8.6%
PRISM-ECW7.2%7.2%6.5%11.3%11.3%
PRISM7.6%6.5%2.0%7.6%7.6%

Here, each policy that omits a dimension leaves it with the largest regret (bold): energy for WaterWise and PRISM-CW, and biodiversity for PRISM-ECW. PRISM bounds all four and has the lowest worst-case regret in all 20 capacity settings. It does so by moving demand from Johannesburg, whose grid has the highest biodiversity intensity, to Tel Aviv.

Together, these experiments separate the two effects. The formulation improves the balance between carbon and water in Figure 7. Including additional dimensions matters when those dimensions would otherwise receive the largest penalty.

Section 7 now describes WaterWise by its weighted objective, introduces PRISM-CW and PRISM-ECW as variants of PRISM, and reports both experiments in Table 2. Appendix A.7.6 gives the per-dimension results (Table 15) and the routing change in the four-region set (Table 16).

AuthorsOpenReview ↗

Response to Reviewer hZ9J (Part 3)

3. Regional hardware differences and network latency

Regional routing evaluation assumes that hardware availability and execution characteristics are identical across all regions. Incorporating regional variations and network latency into the assessment would help further enhance its applicability in real-world scenarios.

Our evaluation includes differences in available capacity across regions, but it holds a configuration’s execution profile fixed across regions. It also does not include network latency. We have clarified these distinctions.

Regional capacity differs in the routing experiments. We generate unequal capacity shares across regions and round them to units of 32 H100 GPUs. We report results across 20 capacity settings, as described in Appendix A.6.2.

However, a given configuration has the same measured energy use and execution time wherever it runs. This isolates how environmental conditions affect placement. Prior geo-distributed work makes the same assumption: WaterWise assumes that datacenters in different regions use the same hardware generation with no performance differences. Deployments with different hardware across regions require region-specific measurements, which PRISM can use directly. They change each request’s impact and resource use in each region, not the formulation.

Network latency requires additional data, particularly the origin of each request and its delay to each candidate region. The Azure trace does not provide request origins, so we cannot evaluate those delays from this trace. We therefore do not claim that the routing results establish end-to-end latency performance for a real deployment. With suitable network data, regions that cannot meet a request’s latency requirement could be excluded from its routing choices. Our existing task-completion-time experiment shows that PRISM can include a performance objective, but it does not replace an evaluation of network latency.

We have clarified the fixed execution-profile assumption and added regional hardware differences and network latency to the limitations in Section 9.

Ending

If these responses resolve your concerns, we would appreciate it if you could reconsider your score. Thank you again for your time and helpful feedback.

Reviewer_zviiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In contrast to prior work which studies the environmental impact of ML models according to a single efficiency cost indicator, the authors investigate the agreement of different dimensions of environmental impact for LLM serving, studying the impacts over energy, water, carbon, and biodiversity. The authors introduce a framework for evaluating the environmental impact of a given workload and serving configuration which determines the settings which minimizes worst-case regret across sustainability dimensions. Concretely, selection of configurations to minimize energy can yield multiple times increases in environmental impact in other dimensions (L394)

Critical strengths
  1. The authors propose a framework for evaluating the environmental impact across sustainability dimensions
  2. The authors identify scenarios in which lowered operational costs on some dimensions but not others lead to disagreement in optimal serving configurations.
  3. The proposed framework is able to reduce regret (maximum environmental impact) across dimensions by over 50%.
Critical weaknesses

Concerns and Questions

  1. Methodological and Writing Clarity.
    • Life cycle assessment terms (e.g. system boundaries and functional units) should be defined for the audience.
    • L237. It is stated that the serving configurations are targeting an identical functional unit. However, that functional unit is not defined explicitly for the chat workload
  2. In Eqn 9, the optimization selects the configuration minimizing the worst-case impact. In practice, is % deviation from the optimal configuration the best measure for evaluating environmental impact. For instance, if a system exists in an energy rich environment where additional power consumption has limited effect on its local surroundings
  3. As impact intensity varies over time, how would the framework account for long running workloads where impact intensities change over the course of execution (e.g. day long training with carbon intensity varying diurnally)?
Questions for authors
  1. Do the authors have intuition as to whether the proposed framework is extensible to other forms of system efficiency or environmental impact beyond those studied? For instance, public health impact [1] or latency?

References

  1. Han, Yuelin, et al. "Health-Informed Computing: Estimating and Addressing the Public Health Impact of Data Centers." arXiv preprint arXiv:2412.06288 (2024).
2 comments
AuthorsOpenReview ↗

Response to Reviewer zvii (Part 2)

4. Workloads that span periods with different impact intensities

As impact intensity varies over time, how would the framework account for long running workloads where impact intensities change over the course of execution (e.g. day long training with carbon intensity varying diurnally)?

PRISM can calculate the impact in each time interval and then add the results. This accounts for changes in both energy use and impact intensity during execution.

For a workload in region rr, the operational impact in dimension mm is Im,op=∑tEIT(t) αm(r,t),I_{m,\mathrm{op}}=\sum_t E_{\mathrm{IT}}(t)\,\alpha_m(r,t), where EIT(t)E_{\mathrm{IT}}(t) is the IT energy used during interval tt, and αm(r,t)\alpha_m(r,t) is the impact per unit of IT energy during that interval. The allocated manufacturing impact is added separately. PRISM then calculates relative regret from the total impacts over the workload. Our experiments test time-varying intensities through hourly serving and routing decisions. The rolling-horizon extension in Appendix A.7.1 also updates decisions as new forecasts become available.

5. Extension to public health and latency

Do the authors have intuition as to whether the proposed framework is extensible to other forms of system efficiency or environmental impact beyond those studied? For instance, public health impact or latency?

Yes. PRISM can include additional impact measures and performance objectives. The required extension depends on how the new objective is measured.

For public health, Han et al. [1] provide a relevant example. Their work estimates health damage from air pollution associated with datacenter electricity use. If health-impact intensities are available, PRISM can combine them with electricity use over location and time. This would require the relevant emissions data and health damage factors. The extension would cover the health effects represented by those inputs.

Task completion time provides a different example because it is measured directly, rather than calculated from energy use. We already evaluate this extension in Section 7 and Appendix A.7.2. PRISM jointly selects configurations and regions while considering task completion time as a fifth objective. In that experiment, it limits the maximum relative regret across completion time and the four environmental objectives to 77.3%.

These examples show that the decision method can compare additional objectives once their values and feasible optima can be calculated. The new objective does not have to follow the same energy-based accounting model. We have added a discussion of the public health extension to Section 9.

References

  • [1] Y. Han, Z. Wu, P. Li, A. Wierman, and S. Ren. Health-informed computing: Estimating and addressing the public health impact of data centers. arXiv:2412.06288, 2024.
  • [2] ISO 14040:2006. Environmental management — Life cycle assessment — Principles and framework. International Organization for Standardization, 2006.
  • [3] U. Gupta et al. ACT: Designing sustainable computer systems with an architectural carbon modeling tool. ISCA, 2022.
  • [4] A. Faiz et al. LLMCarbon: Modeling the end-to-end carbon footprint of large language models. ICLR, 2024.
  • [5] Y. Jiang, R. Kanakagiri, R. Basu Roy, and D. Tiwari. ThirstyFLOPS: Water footprint modeling and analysis toward sustainable HPC systems. SC, 2025.

Ending

If these responses resolve your concerns, we would appreciate it if you could reconsider your score. Thank you again for your time and helpful feedback.

AuthorsOpenReview ↗

Response to Reviewer zvii (Part 1)

Thank you for these helpful comments. We have clarified the lifecycle assessment (LCA) terms, defined the chat workload’s functional unit, and explained the choice of relative regret. We also explain how PRISM handles changes in impact intensity over time and how it can include other objectives. The changes are highlighted in the revised PDF.

1. Definitions of life-cycle assessment terms

Life cycle assessment terms (e.g. system boundaries and functional units) should be defined for the audience.

We have added definitions of these terms in the revision explicitly where they first appear in the revision, following the LCA standard ISO 14040 [2].

  • The system boundary specifies which lifecycle stages the analysis includes. PRISM accounting includes hardware manufacturing and system operation. We now state this boundary when it is first introduced in Section 2. This is consistent with established carbon and water accounting for computing systems, such as ACT [3], LLMCarbon [4], and ThirstyFLOPS [5].
  • The functional unit specifies the service delivered, so that we compare impacts for the same service. For example, we compare the impact of serving one chat request under the same latency and quality requirements. For the agentic workload, we compare the impact per successfully completed task. We now define functional unit at first use in Section 2 and state the workload-specific units in Section 4.

2. Functional unit for the chat workload

L237. It is stated that the serving configurations are targeting an identical functional unit. However, that functional unit is not defined explicitly for the chat workload.

For ShareGPT, the functional unit is one served chat request under latency and quality requirements. Each configuration serves the same workload. A feasible configuration must meet both latency limits: the p90 time to first token (TTFT) must be at most 1,000 ms, and the p90 time per output token (TPOT) must be at most 150 ms. It must also meet the chat-quality threshold specified for the comparison. We measure chat quality through pairwise LLM judging. We have made these units explicit and added references to the service requirements in Section 4.

3. Why use percentage deviation from the optimum?

In Eqn 9, the optimization selects the configuration minimizing the worst-case impact. In practice, is % deviation from the optimal configuration the best measure for evaluating environmental impact? For instance, if a system exists in an energy rich environment where additional power consumption has limited effect on its local surroundings.

Relative regret is useful for comparing objectives with different units, but it does not measure the absolute importance of environmental damage. We have revised the text to make this distinction clear.

In PRISM, relative regret means the percentage increase above the best feasible value for each objective. Equation 9 selects the deployment choice with the smallest maximum percentage increase across the selected objectives. Here, “maximum” refers to the largest loss across objectives, rather than uncertainty about future conditions.

We use this measure because energy, carbon, water, and biodiversity have different units. Their absolute values cannot be compared directly. Relative regret gives each objective a common scale and shows how much a deployment gives up relative to what is feasible for that objective. However, a large percentage increase can still represent a small absolute impact. For example, an increase from 1 to 2 units is a 100% increase, although the absolute increase is only 1 unit. We therefore do not claim that relative regret is the best measure for every setting. We have clarified the purpose and limits of relative regret in Section 3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.