When a large language model answers a question, it does not reread the whole conversation from scratch for every word it writes. It keeps a running memory of what it has already seen, the key-value (KV) cache, and consults that memory for each new word written. Longer contexts and more concurrent users only make the demand for that memory grow, so the bottleneck on modern accelerators is often not compute, but rather how much of this KV cache fits in memory and how fast it can be fetched. Compression can increase capacity and throughput on the same hardware, yet existing solutions have limitations: ordinary floating-point KV barely shrinks under the fast lossless codecs that would suit a memory path, while cutting precision aggressively can save space but may hurt accuracy, often needs custom attention paths, and is hard to revisit once integrated into a serving stack. The goal of this thesis is to develop that middle ground as EMPT-KV: a software-hardware architecture balancing compression effectiveness, generation quality, hardware cost, and long-term feasibility as models and serving stacks evolve. Software applies transforms to the cache to expose structured sparsity that a simple lossless compression codec can exploit, and applies limited in-format precision reduction. Hardware contributes fast decompression on the reload path, restoring ordinary BF16 for attention operations. Across representative models, EMPT-KV provides a 2$\times$ compression ratio while keeping language and task scores near the unmodified BF16 reference. The broader claim is architectural rather than a single codec recipe: transform KV in software keeping the numeric format unchanged, then (de)compress directly on the memory path, easing the KV bottleneck without surrendering quality or future flexibility. ... Read more

Ultrasonic transducers are commonly used in a variety of medical, industrial, and consumer devices, converting electrical energy into high-frequency sound waves. These devices find extensive applications in fields such as medical imaging, non-destructive testing, distance measurement, and cleaning processes. However, these transducers suffer from a narrow operating frequency range caused by a steep frequency response curve with a prominent resonance peak. Existing passive compensation methods using filters are limited due to the individual characteristics of transducers and their susceptibility to process variations, making generic compensation filters impractical. Additionally, the frequency response of transducers changes over time, input power, and environmental conditions, further complicating compensation efforts.

The objective of this project is to overcome these challenges by creating an integrated solution that can provide an ultrasonic transducer system with a consistent frequency response despite external disturbances. The proposed system will incorporate non-linear dynamics characterization and compensation, which are currently lacking in integrated solutions. By accurately characterizing the transducer’s non-linear behavior and compensating for it, the system will overcome the drawbacks associated with passive compensation.

The proposed integrated system holds promising implications for various applications, including medical imaging, material testing, and industrial processes. By mitigating the limitations associated with the narrow operating frequency range of ceramic piezoelectric transducers, this research project contributes to the advancement of ultrasonic technology and its broader impact on diverse industries. ... Read more