Research connecting computing, power and thermal management

Rotary Position Embeddings (RoPE) provide transformers with a fixed grid of positional frequencies, yet trained models use these frequencies highly non-uniformly. We study what determines this frequency usage and propose a data-centered explanation: RoPE frequencies are selected to match the relative-distance structure of the training data. Viewing each frequency as a positional lens, we formalize a field-resolution tradeoff and show that, for a data-induced dependency profile of width $W$, the optimal frequency scales as $1/W$. This frequency-matching principle explains controlled observations on synthetic and text-based data, and suggests that the mid-low frequency bands observed in language models arise from the multi-scale dependency structure of natural language. We further connect frequency selection to position-interpolation-based length generalization: scaling frequencies down expands the effective field while reducing resolution. This helps when longer-context dependencies are approximate dilations of those seen during training, but can fail when relevant dependencies do not scale with context length. Empirically, we show that natural language exhibits approximate self-similarity across positional scales, explaining why test-time frequency scaling can support long-context generalization. Overall, our results identify a data-driven mechanism behind emergent RoPE frequency usage and show that long-context generalization depends on two forms of scale matching: between learned frequencies and training-time dependencies, and between frequency scaling and how those dependencies extend to longer contexts.

Hyperscale data centers and other large concentrated loads can impose substantial new demand on existing transmission networks. If import corridors lack sufficient transfer capability, operators may need to curtail load, delay interconnection, or reinforce the network to maintain reliable service. An energy storage system (ESS) deployed as a storage-as-transmission asset (SATA) offers a non-wires alternative by providing operator-directed support to constrained import corridors. However, the operating-level reliability value of SATA dispatch remains insufficiently quantified. This paper evaluates operator-directed SATA using a day-ahead DC optimal power flow that co-optimizes generation, ESS dispatch, and load curtailment across Monte Carlo scenarios of demand and generator availability. Operating reliability is assessed using expected energy not served (EENS), loss-of-load hours (LOLH), and the conditional value at risk (CVaR) of daily unserved energy. Congestion-price and flow-sensitivity metrics are used to identify the limiting corridor and storage location. The interconnection is then screened to determine whether SATA is suitable, reinforcement is required, or storage would provide little transmission value. Results show that operator-directed SATA reduces average unserved energy, loss-of-load exposure, and tail risk compared with deploying the same ESS for pure arbitrage. These results demonstrate that the operating designation of storage is a primary driver of its transmission value.

Eco-friendly energy management for artificial intelligence data centers (AIDCs) is crucial because of the significant increase in energy consumption-induced carbon emissions from AIDCs resulting from the rapid expansion of AI applications. This paper proposes a hierarchical carbon-aware multi-agent reinforcement learning (CA-MARL) framework for robust and efficient operations of AIDCs under uncertainties while ensuring low-carbon operation of power distribution systems. The framework comprises a workload manager (WM) agent and multiple local AIDC agents trained using a multi-agent transformer method, corresponding to a global AIDC aggregator and a local AIDC operator, respectively. Leveraging AIDC operation data along with nodal carbon intensity (NCI) calculated from the carbon emission flow-integrated distribution system operator problem, the WM agent spatially allocates AI training and inference jobs among all AIDCs. Based on the jobs allocated from the WM agent and NCI information, each AIDC agent schedules economical and eco-friendly operations of the AIDC by performing the following tasks: i) temporal shifting of training jobs, ii) spatial allocation of training graphics processing unit (GPU) blocks and inference GPUs within the AIDC, and iii) control of the supply air temperature of the cooling system. The effectiveness of the proposed framework was assessed using an IEEE 33-node power distribution system.

Photonic Integrated Circuits (PICs) are advancing high-performance computing, data centers, and sensing, yet three-dimensional (3D) PICs introduce critical thermal management challenges due to high-density bonding and heterogeneous materials. Traditional methods like thermal microscopes and in-package sensors yield sparse data, limiting full thermal profile visibility. This paper presents a dual-method solution combining an AI-driven thermal modeling framework with a design-based heuristic approach. The AI method integrates sparse sensor data with design layer and density information to predict multilayer temperature variations, while the heuristic approach uses localized material properties, design layout, component geometries, and sensor coordinates to refine thermal estimations in specific regions. A 2D thermal map of a 3D PIC is generated by interpolating sensor data and adjusting for local thermal resistivity using comparative analysis between design regions. The heuristic method complements the AI model, improving estimation accuracy without extensive training data. Together, these methods offer a scalable, accurate solution for real-time thermal mapping and design-time simulation, enabling reliable thermal management in next-generation 3D photonic systems.

Thermal management in AI data centers is increasingly challenged by bursty workloads and uncertain heat generation. To prevent thermal violations, existing cooling strategies either enforce conservative, rigid bounds that severely limit grid responsiveness, or rely on forecast-driven controllers that perform poorly under AI workload uncertainty and distribution shifts. To overcome the above challenges, this paper proposes a Contextual Distributionally Robust Optimization (CDRO) framework for grid-interactive cooling control. Unlike standard DRO with fixed ambiguity sets, the proposed approach dynamically adapts the Wasserstein radius using real-time AI and grid context. This safely shrinks uncertainty bounds during stable regimes, unlocking deep demand-side flexibility. Theoretically, we formulate the control as an infinite-dimensional inf-sup problem, derive an exact tractable reformulation for the Wasserstein worst-case expected-cost term, and then derive a tractable conservative deterministic counterpart for the Distributionally Robust Conditional Value at Risk (DR-CVaR) thermal safety constraint. Solved via a scalable nested Alternating Direction Method of Multipliers (ADMM) algorithm, the CDRO controller achieves near-zero thermal violations under extreme workload spikes in high-fidelity EnergyPlus co-simulations. Simultaneously, it reduces the operational cost premium of robustness by approximately 13.7 percentage points relative to standard Min-Max Model Predictive Control (MPC).

Large Language Model (LLM) inference workloads are a rapidly growing contributor to data center energy consumption. Optimizing these deployments requires matching specific LLMs to the most efficient GPUs, but operators currently lack the tools to do so without exhaustively profiling each combination. While some predictive models exist, they still require profiling data and struggle to generalize to hardware unseen during training. To address this, we introduce \textit{WattGPU}, featuring two predictive models for mean GPU power draw and Inter-Token Latency (ITL). Our approach leverages only publicly available LLM metadata and GPU specifications, eliminating the need for hardware access or profiling while enabling generalization to unseen NVIDIA server-grade GPUs and LLMs. We evaluate our models using rigorous leave-one-GPU-out and leave-one-LLM-out cross-validation on a dataset of 42 open-source LLMs (0.1B--27B parameters) and 8 GPUs under both offline and server scenarios. The mean power draw model achieves a median absolute percentage error of $\leq3.4\%$ for offline and $\leq13.5\%$ for server scenarios on unseen GPUs, while the latency model achieves $\leq8.5\%$ in server mode, both maintaining strong GPU ranking correlations for server scenarios (Kendall $τ\geq0.76$). Compared to standard physically grounded baselines -- Load-Scaled Thermal Design Power (TDP) for power draw and roofline for latency -- our models reduce median absolute percentage error by approximately 4$\times$ on unseen LLM-GPU combinations for server scenarios or approximately 2$\times$ for completely unseen GPUs. WattGPU's data and code are publicly available at https://github.com/maufadel/wattgpu.

Artificial intelligence depends on large-scale compute resources and their supporting infrastructure. However, AI governance debates treat compute primarily as a technical input rather than as an outcome of investment, ownership, and financial control. This paper examines AI infrastructure investment flows across Africa through a systematic analysis of 46 publicly announced projects totalling USD $12.7 billion between 2019 and 2025. Using a value chain framework, we analyze who invests in AI-relevant infrastructure and where investments concentrate. Our findings reveal a highly concentrated landscape dominated by global data center operators, hyperscale technology firms, and development finance institutions, clustering in South Africa, Kenya, Nigeria, and Egypt. We introduce asymmetrical interdependence to describe a structural condition in which capital and physical infrastructure account for 73% of total funding while control remains concentrated in the compute layer among a small number of global technology firms. We argue that compute governance must account for capital flows, ownership, and control, not only geographic access, because these dynamics shape AI compute equity. Infrastructure presence is necessary but insufficient for meaningful governance capacity.

This paper proposes a grid-aware coordination framework between data centers and distribution grids using a DLMP-based bilevel optimization model. The data center aggregator (DCA) determines active power demand in response to distribution locational marginal prices (DLMPs), while the distribution system operator (DSO) solves a network-constrained optimal power flow problem to determine DLMPs in an unbalanced three-phase system. The model incorporates both active and reactive power consumption of data centers to evaluate their impacts on voltage regulation and phase imbalance. To mitigate adverse network effects, two operating cases are analyzed: without reactive power compensation and with static var generator (SVG)-based compensation. The proposed approach is validated on the IEEE 37-bus unbalanced distribution test system. Simulation results show that DLMP-based coordination captures economically efficient data center operation, and phase- and location-dependent network conditions, while SVG-based compensation improves voltage profiles and reduces phase unbalance.