Research
A Chiplet Expert-Reuse Architecture for MoE Inference
Mixture-of-experts models expand parameter capacity through sparse activation, but Top-k routing creates expert hotspots, random weight loading, inter-chiplet scatter-gather traffic, and synchronization tail latency in chiplet systems.
A Lifecycle-Aware KV Cache System for Agent Inference
Agent inference extends a single model call into a long-lived workflow of planning, tool execution, observation integration, validation, and retries, turning the KV cache from a temporary per-request tensor into execution state spanning multiple turns.
Compiler Optimization for Large Models on Edge NPUs
Edge neural processing units offer energy-efficient matrix arrays, tightly coupled on-chip memory, and asynchronous DMA. However, static graphs, limited dynamic shapes, restricted address spaces, and power and thermal constraints conflict structurally with autoregressive state, variable sequences, and mixed operators in large models.
Portable FP4 Quantization for Heterogeneous Accelerators
FP4 computing for large models is developing into a hardware ecosystem where MXFP4, NVFP4, and vendor-specific variants coexist.
The Evolution of Memory-Centric Architectures for Large-Model Inference
As reasoning models, long contexts, and agent applications evolve, the main inference bottleneck is shifting from insufficient peak compute to frequent movement of model weights, KV caches, expert parameters, and intermediate state across memory hierarchies and interconnects.
Research notes
Define the problem before choosing the model
Research begins not with a model name, but with a question that can be described, tested, and pursued.
Bring equipment data back into operations
The first step in industrial intelligence is not another dashboard, but bringing fragmented field signals back into one decision process.