Autoregressive decoding of large language models becomes bandwidth-limited at long contexts, as generating each token requires streaming the full key-value (KV) cache. We introduce Stochastic Additive No-mulT Attention (SANTA), a method that sparsifies value-cache access by sampling a small number of indices from the post-softmax distribution and aggregating only those value rows via gather-and-add. SANTA yields an unbiased estimator of the attention output while replacing value-stage multiply-accumulates with addition.

In recent years, to reach performance goals, modern computing systems are increasingly turning to using large numbers of compute accelerators, which offer greater power efficiency and thus enable higher performance within a constrained power budget. However, using accelerators increases heterogeneity at multiple levels, including the architecture, resource allocation, competing user needs, and manufacturing variability. Accordingly, current and future systems need to efficiently handle many simultaneous jobs while balancing PM and multiple levels of heterogeneity.

This lecture will discuss Idaho National Laboratory’s mission to, and vision for, accelerating the deployment of AI, research & development, and partnerships that bridge science, engineering, regulation, and industry application. The world is experiencing an energy crisis with rapidly increasing needs and limited supply. The U.S. faces many energy-supply challenges—from access, cost, and infrastructure—as well as dramatic growth in electricity demand driven by hyperscale data centers and electrification.

I will introduce battery materials and microwave synthesis broadly, and the work that former graduate student Kira Wyckoff (supported by an lEE Fellowship) and current graduate student Katie Brockmeyer (supported by an lEE research grant) have carried out. Using domestic microwave ovens, we can quickly convert, in a highly energy efficient manner, inexpensive starting materials into highly effective electrode materials for lithium and sodium ion batteries. I will describe some of the challenges associated with these battery materials and the need for on-site manufacturing.

The past 50 years has seen a dramatic increase in the amount of compute capability per person, in particular, those enabled by AI. It is essential that AI, the twenty-first century’s most important technology, be developed with sustainability in mind. I will highlight key efficiency optimization opportunities for cutting-edge AI technologies. To scale AI sustainably, we must also go beyond efficiency.

Carbon based semiconducting molecular materials now support a wide range of practical technologies, particularly as organic LEDs, OLEDs, used in smartphone and TV displays. The electronic processes that govern their semiconducting properties are strongly controlled by their low dielectric screening, so that excited states, excitons, are often spatially localised and generally show strong magnetic exchange interactions. The exchange interaction presents a challenge for the engineering of efficient OLEDs.

Semiconductor lasers are a key component in numerous emerging photonics-enabled technologies, including additive manufacturing, data centers for artificial intelligence, inertial navigation, optical power beaming, photon sail propulsion for interstellar probes, autonomous vehicles, rapid biomedical diagnostics, laser wakefield acceleration, fusion energy, and directed energy systems. In many of these applications, the laser is the most expensive, power-hungry, and failure-prone component.

The growing energy footprint of Al data centers has stressed local power grids and raised concerns about water stress, air pollution, and derailing climate progress. Analyzing this sector's impacts has proven challenging, given its fast expansion, the rapid evolution of Al hardware, software, and applications, and a pervasive lack of data for accurate energy models.

Computing accelerators, such as GPUs, enhance performance for tasks like high-resolution gaming and Al computations, while specialized accelerators in mobile devices prioritize energy efficiency. Designing these accelerators is complex, as it involves identifying computational tasks best suited for acceleration and optimizing performance and efficiency.