1) Systems Artifact: FastFlowLM
An NPU-first runtime and kernel stack showing practical pathways to low-latency, energy-aware generative inference.
AI Systems and Architecture for Efficient Generative AI
The research agenda focuses on making Generative AI models (LLM, VLM, ASR, Omini, VLA, Diffusion, etc.) efficient across client, edge, and accelerator-rich environments. It combines software runtime design with architecture-level optimization and measurable real-world translation.
Program Components
2) Hardware-Software Co-Design
NPU-native kernels, scheduling, and memory-aware execution strategies that align model behavior with accelerator constraints.
3) Technology Transfer
Integration into AMD platforms and ROCm ecosystem demonstrates movement from university research to production semiconductor software.
Representative Themes
- Memory-centric execution for long-context and generative workloads.
- Quantization- and bandwidth-aware kernel generation for NPUs.
- Runtime systems for predictable latency, throughput, and power envelopes.
- Cross-layer evaluation spanning model behavior, compiler/runtime, and accelerator architecture.
Evidence of Impact
FastFlowLM adoption in AMD's Lemonade Server and migration into ROCm indicate strong practical relevance and reproducible impact beyond publication.
Continue to FastFlowLM Story & Impact for the concise timeline.