projects
Cross-layer heterogeneous memory systems for scalable AI and optimization workloads.
Cross-Layer Heterogeneous Memory Systems for Scalable AI and Optimization Workloads
My research asks how memory systems can better serve AI and optimization workloads through complementary technologies and cross-layer design.
My current focus is GPU and heterogeneous memory-system architecture for AI and LLM workloads.
- AI and optimization workloads can reshape memory-hierarchy requirements.
- Complementary memory technologies can support domain-specific accelerator designs.
- Workload, device, circuit, and system considerations interact, including noise, bandwidth, and thermal tradeoffs.
Architecture project
Cache Level
GPU Register-File Caching and Warp Scheduling
GPU architecture project on register-file caching and warp scheduling
Published work
Domain-Specific Accelerators
Analog eDRAM-RRAM compute-in-memory
Published hybrid-memory work for on-device LLM fine-tuning
Project details
In progress
Main/Storage-Class Memory Level
Exploring high-bandwidth flash within heterogeneous memory systems for GPU-based AI and LLM workloads.
Earlier research
- MAP-SNN publication — Coauthored publication on spiking neural networks
- Charge-domain SRAM compute-in-memory — Charge-domain SRAM-CiM for uncertainty-aware neural networks