arXiv CSopen access

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers Denisa-Andreea Constantinescu1 and David Atienza1

arXiv:2605.26384v1 [cs.DC] 25 May 2026

Embedded Systems Laboratory (ESL), EPFL, Lausanne, Switzerland [email protected]

Abstract. At the global scale, large data centres are growing faster than the electricity grids that supply them. Operators in Europe, North America, and parts of Asia are already discussing rules that will require multi-megawatt computing sites to behave as flexible loads—adjusting their power draw within seconds when the grid needs help absorbing wind and solar. Whether a high-performance computing or artificialintelligence cluster can play that role hinges on a question rarely measured directly: how fast can the software stack take a grid request and actually change the power that the GPUs draw, measured at the facility meter where the grid commitment is settled? We answer this question on real hardware. We build GridPilot, a three-tier predictive controller that adjusts GPU power on three timescales (milliseconds, seconds, hours) and includes a “safety island” that bypasses the slower software path when a fast grid response is needed. On a three-GPU NVIDIA V100 testbed, the measured end-to-end response from a grid trigger to the GPU power crossing its new target is 97.2 ms—6.9 × faster than the strictest European grid product, the Nordic Fast Frequency Reserve. An instantaneous Power Usage Effectiveness (PUE) correction makes the commitment robust at the meter. Experiments against six representative European grids (Sweden through Poland) show that the cooling-overhead drag closed is 2.5–5.8 percentage points (pp). GridPilot is released open source as a proof of concept that MW-scale AI/HPC computing demand can be controllable and flexible by design. Keywords: Power capping · Frequency response · Predictive control · HPC · GPU · PUE.

1

Introduction and Background

1.1

Grid-responsive AI/HPC: context and prior work

Data centres (DCs) are the single fastest-growing source of new electricity demand worldwide; the International Energy Agency projects global DC electricity consumption to roughly double between 2023 and 2030, driven primarily by artificial intelligence (AI) and high-performance computing (HPC) workloads [9]. In parallel, the rising share of variable renewable generation on European, North American, and East Asian grids has eroded the synchronous inertia available to

2

D.-A. Constantinescu and D. Atienza

arrest frequency excursions after disturbances. System operators are responding by procuring an increasingly tight family of frequency-response (FR) services: from Frequency Containment Reserve (FCR, 30 s) down to sub-second products such as the Nordic Fast Frequency Reserve (FFR, 700 ms [16]) and more recent Croatian and Iberian sub-second pilots [20,28]. Multi-MW AI/HPC facilities are being analysed as the most plausible large flexible-load fleet for this market [27,19,25,26]. The prior literature splits into five largely-disjoint streams. (i) Single-layer power-management techniques: the Frontier-scale campaign reports 8.5 % energy savings (1 438 MWh) via system-scale capping [12]; MF-GPOEO reaches 26.2 % via PID-based dynamic voltage and frequency scaling (DVFS) on GPUs [31]; DRLCap attains 22 % via reinforcement learning on H100 boards [32]; Perseus reports training-side capping gains [5]. None of these reports a measured endto-end actuation latency from a transmission-system-operator (TSO)-style trigger. (ii) Hardware-instrumentation work : GPU frequency-switching latencies [29] and the PowerSensor3 precision-measurement harness [30] provide the measurement primitives but are single-layer. (iii) Multi-layer steady-state coordinators: DynamoLLM [23] and hierarchical compute–cooling controllers [1,18] compose multiple control layers but optimise steady-state energy rather than transient response. (iv) Cooling and PUE work : Sun et al. [24] provide the prototype DC cooling decomposition; Zhao et al. [33] apply model-predictive control (MPC) to multi-chiller systems; Liu’s hierarchical rack-to-DC framework [14] attributes 33.8 % of emissions to cooling and reports up to 30 % mean absolute percentage error (MAPE) reduction versus static-PUE accounting when these dynamics are explicit. (v) Holistic in-cluster power-management frameworks: HPE’s PowerSched [22] generalises the Energy Aware Runtime (EAR) [6] and the Global Extensible Open Power Manager (GEOPM) [7] with a system-wide Budget Manager and per-node daemons that allocate power across jobs in an over-provisioned cluster. The EuroHPC JU REGALE library [15] standardises interoperability across these tools through a Data-Distribution-Service (DDS)-based message bus aligned with the HPC PowerStack initiative. These frameworks operate inside a fixed facility power envelope; GridPilot is the orthogonal layer that decides what that envelope should be in response to external grid signals (FFR, Carbon-Free Energy (CFE), demand response, Power-Purchase-Agreement (PPA) matching) — the two compose vertically. The scheduling substrate is the Marconi100 (M100) trace [2], replayed against ENTSO-E [8] hourly carbon intensity (CI) for European grids; CFE is adopted as the user-visible sustainability metric [11]. 1.2

Limitations of the state of practice

The common gap across these five streams is that no prior work, to our knowledge, has measured the end-to-end response of a composed multi-tier controller on a real GPU cluster against a published Transmission System Operator activation budget. First, multiple jurisdictions are preparing rules that will require large AI/HPC DCs to operate as flexible loads over the coming decade; whether

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers

3

AI supercomputers can comply depends on whether the software stack can guarantee deterministic dispatch under realistic workload contention. Second, the most valuable FR products are sub-second [16]; the activation budget is what gates pre-qualification by the TSO. Third, FR commitments are settled at the facility meter, not at the GPU platform level. A controller that ignores instantaneous Power Usage Effectiveness (PUE) under-delivers at the meter by 4–7 pp on warm-water-cooled HPC sites [14]. As large AI-training workloads exhibit 1.5– 2 × higher power swings on H100/H200 hardware [4,17], this latency-and-meter gap will only widen. 1.3

Contributions

This paper presents GridPilot as a proof of concept that MW-scale computing facilities can be designed as flexible, grid-responsive loads. The central engineering claim is small but load-bearing: a deterministic safety-island bypass, composed with predictive tiers above it and a facility-side PUE correction at the top, suffices to meet sub-second budgets at the meter. None of the three is novel in isolation; the contribution is composing them in a way that is measurably correct end-to-end on real silicon. GridPilot makes four contributions. A working prototype of a gridresponsive AI cluster: a three-tier predictive controller (per-GPU at 200 Hz, per-host at 1 Hz, per-cluster hourly) plus an out-of-band safety-island bypass. Measured sub-100 ms grid response on real silicon: on a node with three NVIDIA V100 GPUs, the median end-to-end response is 97.2 ms (101.1 ms worst across 90 trials, at the 700 ms Nordic FFR budget) — an empirical ∼ 6.9× safety margin that is, to our knowledge, the first measured latency under a published TSO budget for a composed multi-tier controller on AI/HPC GPUs. Facilityside accounting via an instantaneous PUE correction: a four-component PUE model added to the cluster-tier objective so the controller’s commitment matches the meter, with replays against six European grids at 1/10/50 MW IT power closing 2.5–5.8 pp of cooling-overhead drag. Open release: the controller, raw V100 telemetry, and reproducibility scripts are released under CC-BY 4.01 .

2

A worked example: what GridPilot does in one second

The Nordic synchronous area requires Fast Frequency Reserve (FFR) providers to deliver their full reserve within 700 ms of a frequency excursion below 49.7 Hz [16]. Picture a 10 MW AI cluster that has pre-committed 2 MW of FFR. At t = 0 a wind plant trips; the grid frequency starts falling. Within one second, GridPilot does the following. (0 ms) The TSO trigger arrives over a dedicated UDP (User Datagram Protocol) socket on the cluster’s control node. 1

https://github.com/denisa-c/gridpilot

4

D.-A. Constantinescu and D. Atienza

(1 ms) The safety island (a small real-time C program pinned to an isolated CPU core) reads the trigger and looks up the new per-GPU power target from a precomputed table. (6 ms) It writes the new power cap to every GPU via the NVIDIA Management Library (NVML); the worst-case NVML cap-update latency is ∼5 ms [29]. (20 ms) The Tier-1 per-GPU proportional-integral-derivative (PID) loop, which runs at 200 Hz independently of the safety island, has already started tracking the new target; the inner-loop step-response settling time we measure is 18–29 ms, depending on workload. (97 ms) The GPU board power crosses 95 % of the new target (median across 90 trials). GridPilot has delivered the reserve. (∼1 s) The Tier-2 per-host predictor, which runs at 1 Hz, sees the new operating point in its next tick and rebalances the per-GPU caps within the host envelope so that the predicted host power one second ahead matches the cluster-tier setpoint. (next hour) The Tier-3 cluster selector picks the next operating point that satisfies the carbon and frequency-response objectives jointly; the PUE-aware correction (Sect. 3.3) makes the IT-side setpoint match the facility-meter commitment. The point of the example is that the latency budget is met because the safety island bypasses the slower predictive layers, not because the predictive layers are themselves fast.

3

GridPilot controller architecture

The architecture of GridPilot, illustrated in Fig. 1, shows the three time-aligned control loops on disparate timescales composed into one pipeline, plus the outof-band safety island used for sub-second response. Electricity grid

Frequency event (TSO trigger)

Tier 3

Tier 2

Tier 1

(hourly)

(1 Hz)

(200 Hz)

Cluster operating-point selector

Per-host coordinator

Per-GPU power-cap loop

Hardware GPU silicon

×✓ ○×

0000

Facility meter

PID Safety island

Real-time C bypass. Reads the grid trigger and writes the GPU cap directly - skipping the slower software path.

Picks IT power target; includes cooling-overhead correction

AR(4) predictor; splits host envelope across GPUs

PID tracking the assigned per-GPU power target via NVML

⏱

Power settles within ~20 ms; meter reflects it ~90 ms after the trigger

real time trigger

Fig. 1. GridPilot architecture. Three control tiers on disparate timescales (per-GPU 200 Hz, per-host 1 Hz, per-cluster hourly). An out-of-band safety island (real-time C, pinned to an isolated core) reads grid triggers and writes GPU caps directly — bypassing the slower software path for real time response.

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers

3.1

5

Three control tiers

Tier 1 — per-GPU PID (200 Hz). The inner loop tracks a per-GPU power target p⋆ via NVML, actuated through nvidia-smi -pl <W>. The control law is the textbook discrete PID form uk = Kp ek + Ki

k X

ei ∆t + Kd

i=0

ek − ek−1 , ∆t

ek = p⋆ − pk ,

(1)

with ∆t = 5 ms, gains (Kp , Ki , Kd ) = (0.6, P 0.05, 0.02) (MF-GPOEO defaults [31], retuned for 200 Hz), anti-windup clamp | e ∆t| ≤ 50 W·s, and saturation at [100 W, 300 W] for the V100 SXM2. The 5 ms tick matches the worst-case NVML cap-update latency [29]. A first-order thermal model with time constant τ = 8 s tracks GPU junction temperature; the inner loop falls back to a 200 W cap if the predicted temperature exceeds 85 °C. Tier 2 — per-host AR(4) coordinator (1 Hz). The outer loop coordinates per-GPU caps within the host envelope through an Autoregressive model of order 4 (AR(4)): the predictor estimates the next sample of host utilisation by running a linear regression on its four most recent past values. Concretely, û(t + 1) =

4 X

αi u(t − i + 1),

(2)

i=1

fitted by Recursive Least Squares (RLS) over a 30 s rolling window (forgetting factor λ = 0.97, 60 s effective memory). Order 4 is chosen by Akaike Information Criterion (AIC) on logged utilisation traces; higher orders give statistically indistinguishable error [29]. Tier 3 — cluster operating-point selector (hourly). The cluster-scale selector performs an hourly grid search over the 2-D space (mean operating fraction µ ∈ {0.4, 0.5, . . . , 0.9}, FR reserve band ρ ∈ {0.0, 0.1, 0.2, 0.3}). The objective is J(µ, ρ) = 0.55 QFFR (µ, ρ) + 0.45 CFE(µ, ρ), (3) following the exogenous-carbon framework of Jahanshahi et al. [10]. QFFR ∈ [0, 1] is the relative FR-provision quality at the facility meter (not at the GPU board); CFE ∈ [0, 1] is the Carbon-Free Energy share [11], the fraction of energy aligned with low-CI windows. Evaluating QFFR at the meter is the requirement that motivates the PUE correction. 3.2

The safety-island bypass

The safety island is the engineering primitive that makes the sub-100 ms latency reproducible. It is an out-of-band component implemented in < 400 source lines of C, statically linked, run as one real-time thread (SCHED_FIFO, priority 80) pinned to an isolated CPU core. A Temporal Logic of Actions (TLA+ ) liveness specification (shipped in the reproducibility kit) proves termination within

6

D.-A. Constantinescu and D. Atienza

four NVML cap-update intervals (20 ms) under the assumption that the kernel honours SCHED_FIFO. The end-to-end budget composes as Le2e = Ltrigger + Ldecide + Lactuate + Lsettle . with Ltrigger ≈ 1 ms (UDP socket read), Ldecide < 50 µs (lookup over the operating-point grid), Lactuate ≈ 5 ms (NVML cap update [29]), and Lsettle ≈ 90 ms (PID settling on the V100 SXM2 thermal plant). The remainder of the 700 ms Nordic FFR budget is reserve. Without the bypass, identical experiments through the Python supervisor stack exhibit p99 dispatch latencies > 250 ms; the median is comparable but the p99 is what would fail TSO pre-qualification. 3.3

PUE-aware FFR control: why facility power matters

European balancing markets settle reserves at the facility meter, not at the GPU board: a controller that commits a 2 MW FFR band on board power may underor over-deliver by ±4–7 pp depending on cooling architecture. Following Sun et al. [24] and the multi-chiller MPC formulation of Zhao et al. [33], we model facility power as PUE(t, L, Tamb ) = 1 +

Pchiller (t) + Ppumps (t) + Pair (t) + Pmisc , PIT (t)

(4)

with L = PIT /PIT,design , pump and air-side affinity laws Ppumps ∝ L2 and Pair ∝ L3 (floored at 20 % and 15 % for bypass-flow and minimum controllability), and a free-cooling fraction ff c (Tamb ) ramping linearly from zero at 25 °C ambient to one at 12 °C wet-bulb. Calibrated to the published Marconi100 design point (PUE = 1.20 at full load), the model achieves up to 30 % MAPE reduction versus static-PUE accounting [14]. A controller that decreases PIT (t) in response to a frequency-restoration request drives PUE(t) up (the L2 /L3 floors bind first), partially offsetting the IT-side swing at the meter. Adding Eq. 4 to the Tier-3 objective closes the gap: the dispatched FFR setpoint matches the metered-side commitment within ±1 pp. The GridPilot carbon- and PUE-aware scheduler operates on hourly dispatch decisions over a 24-hour look-ahead window, integrating four established carbonaware techniques and one new mechanism. Algorithm 1 summarises the dispatch loop. Dynamic power capping during high-CI windows [13] addresses the queueing penalty of pure deferral: capping running jobs delivers carbon savings without adding any wait time. We adopt the 80% power-cap factor from EcoFreq’s recommended default. Composite CI×PUE deferral signal is the new contribution of this work. The dispatch decision uses the product σ(t) = CI(t) × PUE(t, L, Tamb ), normalised over a 24-hour look-ahead window, so that the scheduler defers when σ(t) exceeds the local 66th percentile and dispatches when it does not. This captures

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers

7

Algorithm 1 GridPilot-PUE Dispatch Loop 1: Input: job queue Q, CI signal CI(t), PUE signal PUE(t, L, Tamb ) 2: Compute composite signal σ(t) = CI(t) · PUE(t, L, Tamb ) 3: Compute 24h percentile threshold σ66 4: for each pending job j in priority order do 5: βj ← waitj /dmax,j ▷ aging budget used 6: if σ(t) > σ66 AND βj < 0.7 AND j not short then 7: defer j 8: else 9: dispatch j if nodes available 10: if j is elastic AND j ∈ first 30% then 11: scale replicas inversely to σ(t) 12: end if 13: if σ(t) > σ66 AND j running then 14: apply 80% power cap 15: end if 16: end if 17: end for 18: Backfill short jobs into freed nodes (Extensible Argonne Scheduling sYstem, EASY)

structural carbon savings from two effects: free-cooling alignment (cold-weather wind events that produce low CI also produce low PUE through chiller bypass), and load-dependent PUE (a window with already-high utilisation has lower PUE because misc facility overhead is amortised).

4

Experimental methodology

The methodology answers a single question: given the GridPilot controller, what is the measured end-to-end FR actuation latency on real GPU hardware, and how is it composed across the three tiers and across the four-component PUE model at the facility meter? Testbed. All hardware measurements come from ecocloud-exp06 at EPFL EcoCloud: three NVIDIA Tesla V100 SXM2 32 GB GPUs (300 W thermal design power (TDP) per GPU), 36 physical / 72 logical CPU cores, 379 GiB usable RAM, NVIDIA driver 580.65.06, NVML 12.x, CUDA 12.x, Ubuntu 24.04 LTS. Power-cap actuation uses nvidia-smi -pl <W>, which invokes NVML internally. NVML telemetry is sampled at 100 Hz; CPU and Running Average Power Limit (RAPL) telemetry at 10 Hz. Workload archetypes. Three reference workloads exercise distinct power signatures: matmul (single-stream FP32 general matrix-multiply (GEMM) via cuBLAS Sgemm, pinned near TDP); inference (per-image ResNet-50 batch 1 FP16, memorybound, mean below 200 W); bursty (period-T compute/idle cycle, T = 4 s, 50 % duty).

8

D.-A. Constantinescu and D. Atienza

Experiments. Five instrumented experiments E1–E4, E7 anchor the V100 headline results. E1: power-cap calibration (36-cell power-cap × SM-frequency sweep). E2: inner-loop step response. E3: AR(4) predictor accuracy. E4: closedloop demand-following over 30 s trajectories. E7: end-to-end FR actuation latency, 90 trials (30 per workload, randomised inter-trial delay 5–30 s to defeat caching). A multi-country PUE-aware controller sweep (E8) replays the M100 trace [2] against six hourly CI series at 1, 10, and 50 MW IT power, comparing the CI-only Tier-3 baseline against the PUE-aware variant of Eq. 4. CI is synthesised from European Environment Agency (EEA) and Ember country means and the 2020–2024 ENTSO-E [8] diurnal envelope; the released kit also ships a real-CI fetcher (ENTSO-E A75 Actual Generation per Production Type with IPCC AR5 lifecycle factors). A supervisory cross-tier experiment (E5) ships in the kit as a design only — the three-GPU testbed cannot sustain the multi-day operation it would need. Metrics. Iterations-per-joule (energy efficiency at the calibrated operating point); AR(4) one-step-ahead Mean Absolute Error (MAE, W); closed-loop tracking error (percentage point [pp] vs. host-envelope setpoint); end-to-end FR actuation latency (ms, UDP-trigger to NVML-cap-setpoint-reached at 95 % of target); ∆facility (percentage IT-side vs. facility-side CO2 reduction at constant CFE); and Net CO2 = Operational CO2 − Exogenous CO2 , capturing both workloadside emissions (energy × grid CI) and avoided reserve-side emissions from provided FFR, so PUE-aware dispatch can be assessed when lower operational CO2 coincides with reduced high-CI FFR provision.

5

Measured results

5.1

Component-level validation

Power-cap calibration (E1). The 36-cell power-cap × streaming-multiprocessor (SM) frequency sweep yields a best-efficiency operating point of pcap = 150 W and fsm = 945 MHz across all three workloads, within ±5 % on iterationsper-joule. Best iterations-per-joule values are 2.880 / 0.570 / 0.549 for inference / matmul / bursty. The same sweep fits the per-workload power model PGPU = Pidle +αf +βf 2 L+γL with Pidle = 39 W; leave-one-out cross-validation MAE is 3.45 %, and the 980-node-scale envelope matches the published Marconi100 facility-power reference within +1.4 %. Inner-loop step response (E2). A step command p⋆ : 280 → 200 W at t = 0, logged at 100 Hz, settles to within ±2 % of the new setpoint in a median of 18 / 21 / 29 ms (matmul / inference / bursty).; Fig. 2 shows a representative trace. This validates the 200 Hz Tier-1 cadence and indicates that the 700 ms FR budget is dominated by the dispatch path, not by actuator settling. Outer-loop AR(4) predictor accuracy (E3) Fig. 3 (left) reports the perworkload MAE of the 1 Hz predictor over 50–66 one-step-ahead predictions on a 30 s rolling window. Inference (4.69 W) is the tightest because of its nearstationarity; matmul (7.00 W) is moderately higher because of GEMM tile-

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers E2 step response

pcap=280W (initial high)

GPU power (W)

250

E2_inner_loop

9

GPU 0

pcap=200W (step-down) pcap=280W (step-up)

200 150 100 50

0

10

20

Time (s)

30

40

Fig. 2. Inner-loop step response on the V100 testbed: step-down from 280 W to 200 W and return step-up, with rapid settling inside the control band.

schedule variance; bursty (19.66 W, ∼ 3× matmul) is bimodal at the 30 s window. The bursty p95 envelope is the residual that the cascade absorbs at Tier-2.

164.1

100 50 4.69 12.8 0

inference

7.00

20.0

matmul

19.66 bursty

14 12

E4 — Demand tracking error 5% threshold

10 8 6 4 2 0

1.68% (3.5 W) inference

2.12% (4.4 W)

matmul

E7 — Per-trial FFR latency

105.0

11.08% (23.0 W)

bursty

Margin

700

700

102.5

600

100.0

500

97.5

400

ms

MAE p95

Latency (ms)

150

E3 — AR(4) predictor accuracy Relative MAE (%)

Prediction error (W)

200

~7×

300

95.0

200

92.5

100

90.0

0

inference matmul

bursty

101

maxbudget

Fig. 3. V100 hardware results. (a) AR(4) one-step-ahead MAE per workload (4.69 / 7.00 / 19.66 W for inference / matmul / bursty). (b) Closed-loop demand-following tracking error; the 5 % band is the cascade-composition diagnostic, not a failure mode. (c) End-to-end FR actuation latency over 90 trials (median ∼97.2 ms; max 101.1 ms; 90/90 pass at the 700 ms Nordic FFR budget).

Closed-loop demand-following (E4). A 30 s demand-following trajectory tests the composed Tier-1+Tier-2 cascade. Inference (1.68 %) and matmul (2.12 %) fall within the 5 % acceptance band; bursty (11.08 %) exceeds it, consistent with the outer-loop bursty p95 envelope reported below. The 5 % threshold is not a failure mode but a cascade-composition diagnostic: below it the Tier-1 PID tracks alone, above it the Tier-2 host predictor absorbs the residual. End-to-end FR actuation latency (E7). Across 90 trials the median end-toend latency is 97.2 ms (matmul), 97.5 ms (inference) and 97.8 ms (bursty); the maximum across all 90 trials is 101.1 ms; the 90/90 pass rate at the 700 ms Nordic FFR budget gives an empirical safety margin of ∼ 6.9×. Fig. 3 (right) shows the distribution. Without the safety-island bypass, identical experiments through the Python supervisor stack exhibit p99 latencies exceeding 250 ms (garbage-collection pauses and lazy-import blocking on first call). End-to-end multi-tier controller validation. Fig. 4 extends the validation to cluster scale: a 24-hour simulation of a 100-host cluster on the German grid reproduces AR(4) accuracy (MAE 0.036, p95 = 0.09) and per-GPU tracking (mean 102 W, p95 396 W) at perfect (1.0) FFR-provision quality with a 20 % reserve

D.-A. Constantinescu and D. Atienza

0.8

380

Carbon intensity (gCO₂/kWh)

Operating fraction

Mean op fraction FFR reservation

360 340

0.6

320 0.4

300 280

0.2

0.0 0

260 240 10

Hour (UTC)

(b) AR(4) predictor MAE=0.036, p95=0.090

400

(c) CFE alignment by grid (median, [diurnal range]) 100

y = x ideal

0.8 0.6 0.4 0.2 0.0 0.00 0.25 0.50 0.75 1.00

Actual utilisation

20

50% [28-78]

70%

1.0

AR(4) predicted

1.0

80 [65-75]

60 30% [22-40]

40

20

0

CH

IT

DE

(d) Savings decomposition Operational (CFE) Exogenous (FFR) 26%

30

Net CO₂ savings (%)

(a) Tier-3 op-point trajectory

CFE alignment (%)

10

25 21%

20

3%

20%

8%

6%

15 10

18%

18% 14%

5 0

CH

IT

DE

Fig. 4. Multiscale controller validation. (a) Tier-3 operating-point trajectory on the German grid over 24 hours. (b) Tier-2 AR(4) predictor fit on host utilisation. (c) Carbon-free-energy alignment across representative grids. (d) Net-savings decomposition into operational and exogenous components at 50 MW scale.

band, with operating-point selection at 0.90 mean utilisation in green-rich daytime windows versus 0.40 overnight. At 50 MW scale, the net-savings decomposition for CH/IT/DE totals 21/20/26 % (DE shows an 8 % exogenous contribution); the simulator runs at over 26,000× real-time, supporting digital-twin replay and real-time deployment. 5.2

Multi-country PUE-aware controller sweep (E8)

Fig. 5 reports the cooling-overhead drag closed by the PUE-aware Tier-3 selector versus the CI-only baseline across six European grids at 10 MW IT power (panel a), and across 1/10/50 MW for the Sweden (cleanest) and Poland (dirtiest) bookends (panel b). The envelope is widest on low-CI grids — where IT-side savings are small in absolute terms and the cooling-overhead becomes a larger fraction of facility power — and narrowest on high-CI grids. Cluster-scale averaging shrinks the envelope at 50 MW because the L2 /L3 cooling floors bind less often. The 50 MW envelope of 2.5–5.8 pp is the operationally relevant range for European deployments at the Marconi100 design-point PUE of 1.20; it scales roughly linearly with (PUEdesign − 1), so warm-water sites at PUE ≈ 1.10 see proportionally smaller gains and chilled-water hyperscale at PUE ≈ 1.30 proportionally larger. A full PUEdesign sensitivity sweep (E9, PUEdesign ∈ [1.10, 1.40]) is planned as journal-extension follow-up.

6

Discussion and limitations

The experimental campaign uses three V100 SXM2 GPUs at a single site. Rackscale contention and warm-water cooling at 10–50 MW are projections, not measurements. H100/H200/MI300 architectures exhibit 1.5–2× higher power swings [4,17] and the cascade has not yet been validated on those at the FRlatency layer. The synthetic TSO trigger used in E7 is not a TSO pre-qualification;

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers (a) Cooling-overhead drag closed at 10 MW IT

(b) Cluster-scale averaging shrinks the envelope

50

SE (CI=11 g/kWh) PL (CI=612 g/kWh)

50

40

Δfacility (pp)

Δfacility (pp) from PUE-aware FFR

60

30

40 30

20

20

10

10

0

11

SE 11

CH 30

FR 53

IT 258

DE 295

PL 612

0

1 MW

10 MW

50 MW

Country (annual mean CI g/kWh below)

Fig. 5. PUE-aware FFR controller. (a) ∆facility (percentage points) at 10 MW IT, one bar per country, ordered by mean CI. (b) MW scaling for the SE and PL bookends.

full pre-qualification requires integration with PICASSO [3] or MARI [20]. A supervisory cross-tier experiment (E5) ships in the GridPilot kit as a design only. Lessons learned. L1: the 5 % tracking threshold is a diagnostic (bursty hits 11.08 %), not a failure mode. L2: the ∼97.2 ms FR latency is reproducible only with the deterministic safety-island bypass — Python-only implementations produce p99 > 250 ms. L3: facility-meter accounting is the binding correctness criterion — a controller that ignores the four-component PUE under-delivers at the meter by 4–7 pp.

7

Conclusion and Outlook

GridPilot is a proof of concept that a real GPU cluster can be made gridresponsive at the sub-second scale that European FR products demand of MWclass AI/HPC data centres. The engineering claim is two-part: a deterministic safety-island bypass meets the strictest published activation budget on commodity GPU hardware with a measured ∼ 6.9× margin, and a four-component PUE correction closes 2.5–5.8 pp of cooling-overhead drag at the facility meter across six European grids. GridPilot is the grid-facing layer of the emerging HPC PowerStack, composing vertically with in-cluster power managers (PowerSched [22], EAR [6], GEOPM [7]) through the REGALE DDS bus [15] and aligning with SEANERGYS [21]. The kit is open-source and reproduces the V100 campaign in ≤ 48 GPU-hours. Natural next steps are porting to H100/H200/MI300 platforms, where larger power swings make the multiscale design more critical, and a production prototype on a PowerSched-enabled over-provisioned system that would consume GridPilot’s Tier-3 setpoint over the REGALE message bus. Reproducibility release. The GridPilot v1.0 kit ships under MIT (code) and CC-BY 4.0 (data) at https://github.com/denisa-c/gridpilot. the complete controller stack, the safety-island C skeleton with TLA+ specification, the raw 100 Hz NVML telemetry for every reported V100 experiment, the per-workload power-model calibration scripts and residuals, the multi-country sweep driver, the figure-regeneration pipeline, the editable architecture deck and a pytest suite

12

D.-A. Constantinescu and D. Atienza

covering controller invariants and data-loader contracts. A step-by-step measurement protocol reproduces the full campaign in ≤ 48 GPU-hours on a comparable 3× V100 testbed.

Acknowledgements This work has been partially supported by the EPFL Solutions 4 Sustainability program “HeatingBits: renewable-supplied data centers integrating heating and cooling supply of local districts” and the UrbanTwin project (ETH Board Joint Initiatives for the Strategic Area Energy, Climate and Environmental Sustainability, and the Strategic Area Engagement and Dialogue with Society). The authors also thank the EcoCloud center of EPFL, in particular Dr. Xavier Ouvrard, for providing access to the V100 server node.

References 1. Abera, N.B., et al.: Coordinated cooling and compute management for AI datacenters (2025), https://arxiv.org/abs/2511.08123 2. Antici, F., Seyedkazemi Ardebili, M., Bartolini, A., Kiziltan, Z.: PM100: A job power consumption dataset of a large-scale production HPC system. In: Proceedings of the SC ’23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis. ACM (2023). https://doi.org/10. 1145/3624062.3624263 3. Backer, M., Kraft, E., Keles, D.: The economic impacts of integrating european balancing markets: The case of the newly installed afrr energy market-coupling platform PICASSO. Energy Economics 128 (2023). https://doi.org/10.1016/ j.eneco.2023.107095 4. Choukse, E., Warrier, B., Heath, S., Belmont, L., Zhao, A., Khan, H.A., Harry, B., Kappel, M., et al.: Power stabilization for AI training datacenters (2025), https: //arxiv.org/abs/2508.14318 5. Chung, J.W., Gu, Y., Jang, I., Meng, L., Bansal, N., Chowdhury, M.: Perseus: Reducing energy bloat in large model training. In: Proceedings of the ACM SIGOPS 30th Symposium on Operating Systems Principles (SOSP ’24). ACM (2024). https://doi.org/10.1145/3694715.3695970 6. Corbalán, J., Vidal, O., Casas, M., Alonso, D.: EAR: Energy management framework for supercomputers. Technical report, Barcelona Supercomputing Center (2020) 7. Eastep, J., Sylvester, S., Cantalupo, C., Geltz, B., Ardanaz, F., Al-Rawi, A., Livingston, K., Keceli, F., Maiterth, M., Jana, S.: GEOPM: A scalable open runtime framework for power management. In: Proceedings of the International Supercomputing Conference (ISC) (2017). https://doi.org/10.1007/978-3-319-58667-0_ 21 8. ENTSO-E: ENTSO-E transparency platform. ENTSO-E Transparency Platform (2015), https://transparency.entsoe.eu 9. International Energy Agency: Electricity 2025: Analysis and forecast to 2030. IEA Report, Paris (2025), https://www.iea.org/reports/electricity-2025

GridPilot: Real-Time Grid-Responsive Control for AI Supercomputers

13

10. Jahanshahi, A., et al.: Coordinating power grid frequency regulation service with data center load flexibility (ecocenter) (2025), https://arxiv.org/abs/2511. 05721 11. Kamatar, A., Gonthier, M., Hayot-Sasson, V., Bauer, A., Copik, M., Hoefler, T., Castro Fernandez, R., Chard, K., Foster, I.: Core hours and carbon credits: Incentivizing sustainability in HPC (2025), https://arxiv.org/abs/2501.09557 12. Karimi, A.M., Maiterth, M., Shin, W., Sattar, N.S., Lu, H., Wang, F.: Exploring the frontiers of energy efficiency using power management at system scale. In: SC24W: Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis. IEEE (2024) 13. Kozlov, O., Stamatakis, A.: Ecofreq: Compute with cheaper, cleaner energy via carbon-aware power scaling (2024), https://arxiv.org/abs/2410.01533 14. Liu, W.: Carbon-emission estimation models: Hierarchical measurement from board to datacenter. Journal of Industrial Engineering and Applied Science (2026) 15. Madella, G., et al.: The REGALE library: A DDS interoperability layer for the HPC PowerStack. Journal of Low Power Electronics and Applications (2025) 16. Manner, P., Tikka, V., Honkapuro, S., Tikkanen, K., Aghaei, J.: Electric vehicle charging as a source of Nordic fast frequency reserve — proof of concept. IET Generation, Transmission & Distribution (2023). https://doi.org/10.1049/ gtd2.13042 17. Newkirk, A.C., Fernandez, J., Koomey, J., Latif, I., Strubell, E., Shehabi, A., Samaras, C.: Empirically-calibrated H100 node power models for accurate AI training energy estimation. Environmental Research: Energy 2(4) (2025). https: //doi.org/10.1088/2753-3751/ae2486 18. Ottaviano, A., Bambini, G., Tortorella, Y., et al.: ControlPULP: A risc-v on-chip parallel power controller for many-core hpc processors with hardware/software realtime control. In: International Journal of Parallel Programming (2023). https: //doi.org/10.1007/s10766-023-00761-w 19. Ren, P., Sun, W., Wang, Y., Harrison, G.: Grid frequency stability support potential of data center: A quantitative assessment of flexibility (2025), https: //arxiv.org/abs/2510.01050 20. Sagrestano Štambuk, P., Vrbičić Tenđera, D., Zovko, N., Tenđera, T., Uzelac, M.: Alignment of aFRR and mFRR prequalification process in Croatia with the target market design. Journal of Energy – Energija 72(3), 3–7 (2023). https://doi.org/ 10.37798/2023723472 21. SEANERGYS Consortium: SEANERGYS: Software for efficient and energy-aware supercomputers. EuroHPC JU HORIZON-EUROHPC-JU-2023-ENERGY04 (2025), https://www.eurohpc-ju.europa.eu/research-innovation/ our-projects/seanergys_en 22. Simmendinger, C., Marquardt, M., Mäder, J., Schiffmann, T., Wilde, T.: PowerSched – managing power consumption in overprovisioned systems. In: 2024 IEEE International Conference on Cluster Computing Workshops (CLUSTER Workshops). IEEE (2024) 23. Stojkovic, J., Zhang, C., Goiri, Í., Torrellas, J., Choukse, E.: DynamoLLM: Designing LLM inference clusters for performance and energy efficiency. In: Proceedings of the 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE (2024) 24. Sun, K., Luo, N., Luo, X., Hong, T.: Prototype energy models for data centers. Energy and Buildings 231 (2020). https://doi.org/10.1016/j.enbuild.2020. 110166

14

D.-A. Constantinescu and D. Atienza

25. Takçı, M.T., Qadrdan, M., Summers, J., Gustafsson, J.: Data centres as a source of flexibility for power systems. Energy Reports 13 (2025). https://doi.org/10. 1016/j.egyr.2025.04.013 26. Tao, X., Gadh, R.: Fast frequency response potential of data centers through workload modulation and UPS coordination. IEEE Access 13, 145110–145125 (2025). https://doi.org/10.1109/ACCESS.2025.3646120 27. Terzija, V., et al.: Data centers for sustainable grids: From microgrids to supergrids. IEEE Energy Sustainability Magazine (2026), https://ieeexplore.ieee. org/document/11367124/ 28. Varhegyi, G., Nour, M.: Integrating fast frequency response ancillary services: A global review of technical, procurement, and market integration challenges. Clean Energy 9(2), 204–218 (2025). https://doi.org/10.1093/ce/zkae064 29. Velicka, D., Vysocky, O., Říha, L.: Methodology for GPU frequency switching latency measurement. In: 2025 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW). pp. 830–839. IEEE (2025). https: //doi.org/10.1109/IPDPSW66978.2025.00133 30. van der Vlugt, S., Oostrum, L., Schoonderbeek, G., van Werkhoven, B., Veenboer, B., Doekemeijer, K., Romein, J.W.: PowerSensor3: A fast and accurate open source power measurement tool. In: Proceedings of the 2025 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). IEEE (2025) 31. Wang, F., Hao, M., Zhang, W., Wang, Z.: Model-free GPU online energy optimization. IEEE Transactions on Sustainable Computing 9(2), 128–141 (2024) 32. Wang, Y., et al.: DRLCAP: Runtime GPU frequency capping with deep reinforcement learning. IEEE Transactions on Sustainable Computing (2024) 33. Zhao, J., Chen, Z.x., Li, H., Liu, D.: A model predictive control for a multi-chiller system in data center considering whole system energy conservation. Energy and Buildings (2024). https://doi.org/10.1016/j.enbuild.2024.114919