Session: ISMM - ISMM: Keynote 2 - Dr. Euicheol Lim
summary
Shows that mainstream LLMs drive demand for HBM and expose GPU bandwidth and latency limits using ChatGPT as an example, arguing current GPU-based stacks hit a data‑movement and cost bottleneck. Presents a working prototype of AiM memory device paired with an AiMX LLM accelerator that uses processing‑in‑memory to cut memory traffic and inference cost, and outlines end-to-end memory+compute fusion architectures for running real AI services.
AI-written summary · about
related talks
23mFrom Hallucinations To Hardware: Diagnosing LLM Failures
14mSYMPHONY: Enabling Compute-Memory Disaggregation in LLM Serving Systems
28mSimplifying AI for Edge Compute with HAMi
13mKeynote: How We Save $900 per Day with Self-Hosted AI: Building Scalable Local LLM Infrastructure
01h 18mOptimizing LLM Training and Inference Performance on GPUs
11mEngineering Scalable ML Infrastructure for Modern Apps
27mShattering the AI Memory Wall
26mEfficiently serving LLMs at scale.