HYDRA preprint reports faster hybrid-LLM serving on chiplet systems

A new arXiv paper describes HYDRA, a design-space exploration framework for serving hybrid Transformer-Mamba language models on heterogeneous chiplet systems. The framework jointly evaluates chiplet composition and placement, inter-chiplet bandwidth, batching and runtime scheduling. Across its workloads, the paper reports average throughput 1.55 times higher and time-to-first-token 43.7% lower than state-of-the-art baselines, with throughput gains up to 2.3 times. The results are presented as performance comparisons from the framework's evaluation.