Analog photonic multiply-accumulate (MAC) processors are a promising alternative to digital electronics for energy-intensive workloads such as artificial intelligence. Prototypes include photonic crossbar arrays, Mach–Zehnder interferometer networks, and microring resonators. However, predicting large-scale performance remains challenging due to inconsistent benchmarking of system-level energy metrics. We address this gap by evaluating the energy budget and scaling behavior of analog photonic processors. Our framework establishes a unified benchmarking approach, enabling objective comparison across architectures and guiding the development of energy-efficient photonic hardware. Any analog photonic MAC processor integrated with conventional compute requires both electro-optical and digital-to-analog conversions, and vice versa. At scale, the per-MAC costs of digital and analog conversions decrease linearly with the matrix size, while costs of electro-optical conversions depend on architecture and associated propagation losses. We show that low-loss optical circuits are essential to surpass state-of-the-art electronic hardware in energy efficiency. Finally, we propose a scalable hybrid electro-optical MAC processor that reduces optical losses while operating with incoherent light, providing a path toward reprogrammable energy-efficient photonic accelerators tailored for AI workloads.
Floating-point precision tuning is essential for optimizing numerical software, involving the selection of the ideal precision for floating-point variables to balance performance and accuracy. Identifying the optimal configuration is challenging due to the large search space and the interactions between precision and speed. Existing works get stuck in local optima, removing a variable and/or a function call from the search space immediately if it fails to meet a given error threshold. We introduce RLTuner, a novel automated tuning framework that reformulates precision tuning as a Markov Decision Process (MDP) and employs a Deep Q-Network (DQN) to identify superior configurations that traditional pruning-based methods overlook. We evaluate RLTuner on 20 numerical benchmarks, ranging from kernels to large-scale applications like LULESH. RLTuner outperforms state-of-the-art baselines in 17 out of 20 cases, achieving speedups of up to 46.28%, and yields the lowest error in 11 out of 20 cases with up to 10 orders of magnitude higher accuracy. Notably, it identifies higher-quality configurations—yielding both superior performance and lower error—in 40% of benchmarks.
As exascale systems increasingly rely on hardware overprovisioning to sustain throughput under strict power limits, accurate, fine-grained power prediction becomes critical. We present a machine learning framework leveraging lightweight, always-on NVIDIA DCGM telemetry from the Perlmutter super-computer to predict runtime GPU power consumption, designed with real-time deployment constraints in mind. Using a month-long production telemetry dataset for training and evaluation, we demonstrate that a generic, application-agnostic model achieves an average Root Mean Squared Error (RMSE) of 50 W. Specialized application-aware models for dominant workloads (e.g., Chroma, VASP) further reduce prediction error by up to 48% compared to generic models, while combining DCGM metrics with historical power data improves accuracy by up to 15.8% over power-only models with negligible inference latency. Our framework provides a scalable building block for dynamic power management systems that allow hardware-overprovisioned systems to reclaim performance otherwise lost to static power cans.
Kinetic Monte Carlo is a cornerstone for rare-event dynamics in materials, but its strict sequentiality and super-basin trapping create a scalability deadlock: massive computation advances physical time only marginally, throttling both spatial and temporal scalability. We present EscapeKMC, the first reinforcement learning-guided KMC framework that resolves the scalability dead-lock under strict physical alignment. EscapeKMC introduces three techniques: (1) Poisson Clock Alignment anchors event timing to the Poisson law and enables $O(1)$ escape-time estimation; (2) Adaptive Swarm Reasoning distributes decision-making across lightweight atomic agents coordinated by a centralized critic, yielding system-size-invariant policies transferable across scales; and (3) Sparse Dynamics Regularization restructures irregular dynamics into batched, accelerator-friendly execution efficiently. On reactor-pressure-vessel benchmarks spanning concentrations from 10-5 to 10-1 and temperatures from 663–773 K, EscapeKMC achieves $17.5\times$ over OpenKMC while maintaining physical fidelity. At extreme scale, it simulates $5.8\times 10^{10}$ atoms on a single NVIDIA A100 GPU with a $500\times$ improvement in memory efficiency. Beyond single-GPU execution, EscapeKMC scales efficiently to large GPU clusters. It sustains 91% parallel efficiency on up to 16,384 GPUs and enables KMC simulations of up to $1.7\times 10^{14}$ atoms, pushing atomistic irradiation modeling to previously unattainable scales.
As we make strides toward the era of quantum utility with improved hardware and defined practical use cases, a primary challenge resides in the definition of robust implementation architectures that enable efficient utilization of quantum accelerators. To truly unlock the capabilities of Quantum Processing Units (QPUs), basic classical support must be upgraded to a high-performance, efficient infrastructure. On the application level, integration translates to defining coherent workflows that enable optimized implementations, while abstracting complexity for larger tasks. Accordingly, a common primitive in hybrid applications is the evaluation of a physical system's observable quantities. Leveraging quantum hardware for this task requires decomposing observable-derived operators into vast arrays of Pauli terms, necessitating a high volume of distinct circuit executions that currently suffer from suboptimal management. This work proposes a novel, scalable procedural architecture that establishes observable measurement as a first-class citizen of the High-Performance Computing - Quantum Computing (HPCQC) stack. We introduce a hierarchical runtime environment featuring a workload manager for resource-aware scheduling and local quantum and classical workers for low-latency, device-specific interactions. Such infrastructure enables a pipeline of advanced optimizations, including, but not restricted to, commutativity-based operator grouping, adaptive shot allocation for variance-aware load balancing, and stateful circuit caching to minimize compilation overhead. By automating the orchestration of measuring non-commuting terms across heterogeneous backends, we demonstrate a pathway to significantly reduce time-to-solution and improve accuracy. This procedure bridges the gap between theoretical quantum algorithms and practical, scalable HPC deployment.
