News

05/26: uAgent, 7nm edge chip for LLM reasoning with adaptive block-wise mixed-format 4-bit quantization, accepted at VLSI.

02/26: GaussianVision accepted at CVPR! In this work, we demonstrate that 2D Gaussian Splatting can serve as an efficient and effective visual substrate for large scale vision–language alignment while providing 3×–23.5× compression and data-loading speedups.

02/26: AgenticCache accepted at MLSys! In this work, we propose a cache-driven asynchronous planning framework that decouples LLM reasoning from real-time execution, reducing agentic latency by 75%. 

02/26: Alpha-Vision, 16nm always-on vision processor with 787us face detection latency in <5mW, presented at ISSCC.

09/25: We demonstrate a new compressed representation substrate for vision-language alignment using 2D Gaussian Splatting. Link. 

05/25: Congratulations to Yasmine Omri on being selected a 2025 Knight-Hennessy Scholar! 

05/25: BlockDialect, our 4-bit mixed- number format codebook encoding for efficient LLM computation is accepted at ICML'25!

04/25: GainSight, our data lifetime profiling infrastructure for aiding the composition of heterogeneous on-chip memories in AI accelerators is open-source. 

04/25: Our work on token sequence compression for efficient multimodal computing is out. 

02/25: BlockDialect, our 4-bit mixed- number format codebook encoding for efficient LLM arithmetic is open-source.