Schedule
| Week | Lecture 1 | Lecture 2 |
|---|---|---|
| Jan 14/16 | Introduction to Deep Learning Systems; GPU/Accelerator Overview [Slides] | S26-Week-1-Lect2-CIFAR10.pptx — CIFAR-10 example and basic DL workflow |
| Jan 21/23 | Modern GPU Architecture for Deep Learning | S26-Week-2-Lect2-GEMM.pptx — Matrix multiplication and GPU compute cores |
| Jan 28/30 | CUDA Programming and Kernel Optimization | S26-Week-3-Lect2-TensorCore.pptx — Tensor Core programming and mixed precision |
| Feb 4/6 | Memory Hierarchy, Caching, and Memory Management | S26-Week-4-Lect2-Convolution.pptx — Convolution kernel optimization and dataflow |
| Feb 11/13 | Tensor Operations and Optimized Kernels | S26-Week-5-Lect2-Transformer.pptx — Transformer model and attention computation |
| Feb 18/20 | Compiler Techniques for Deep Learning (IR, Operator Fusion) | S26-Week-6-Lect2-SparseMM.pptx — Sparse matrix multiplication and graph IR |
| Feb 25/27 | Distributed Training Fundamentals; Data vs Model Parallelism | S26-Week-8-Lect2-DistTrain.pptx — Distributed training implementation |
| Mar 4/6 | Communication Optimizations and Scheduling | S26-Week-9-Lect2-Diffusion.pptx — Diffusion models and compute scheduling |
| Mar 11/13 | Systems for Large Models (LLMs, Diffusion Models) | S26-Week-10-Lect2-DLRM.pptx — Deep Learning Recommendation Models |
| Mar 18/20 | Profiling, Benchmarking, and Performance Analysis | (TBD — profiling demo or tool lecture) |
| Mar 25/27 | Case Studies: GNN Acceleration, Diffusion Models, Recommender Systems | (Combine with prior GNN, Diffusion, and DLRM lectures as recap) |
| - Apr 24 | Course Presentation | Course Presentation |
| Date | Lecturer | Topic | Materials |
|---|---|---|---|
| Feb 18 | Jianming TongBioJianming Tong (https://jianmingtong.github.io/) is a 4th-year PhD candidate at Georgia Tech, a visiting researcher at MIT. He focuses on full-stack optimizations—spanning model, system, compiler, and hardware—for enhancing both efficiency and privacy of AI systems. He proposed a framework to approximate non-linear ML operators as polynomials to be compatible with Homomorphic Encryption (HE) without utility sacrifice, enabling privacy-preserving ML via HE (model, MLSys'23), and developed the CROSS compiler to convert HE workloads as AI workloads to be accelerated by existing Google TPUs, enabling immediate scalable low-cost privacy-preserving capability to existing AI stacks and designed a dataflow-layout co-switching reconfigurable accelerator for efficient inference of dynamic AI workloads (ISCA'24). These works are widely deployed in NVIDIA, Google, IBM, and recognized by Qualcomm Innovation Fellowship and Machine Learning and System Rising Star.Lecture AbstractThe inference efficiency of diverse ML models over spatial accelerators boils down to the execution of different dataflows (i.e. different tiling, ordering, parallelism, and shapes). Using the optimal dataflow for every layer of workload can reduce latency by up to two orders of magnitude over a suboptimal dataflow. Unfortunately, reconfiguring hardware for different dataflows involves on-chip data layout reordering and datapath reconfigurations, leading to non-trivial overhead that hinders ML accelerators from exploiting different dataflows, resulting in suboptimal performance. To address this challenge, we propose FEATHER, an innovative accelerator that leverages a novel spatial array termed NEST and a novel multi-stage reduction network called BIRRD for performing flexible data reduction with layout reordering under the hood, enabling seamless switching between optimal dataflows with negligible latency and resources overhead. For systematically evaluating the performance interaction between dataflows and layouts, we enhance Timeloop, a state-of-the-art dataflow cost modeling and search framework, with layout assessment capabilities, and term it as Layoutloop. We model FEATHER into Layoutloop and also deploy FEATHER end-to-end on the edge ZCU104 FPGA. FEATHER delivers 1.27 | FEATHER | [TBD] |
| Feb 18 | Hongzhen Chen | [TBD] | [TBD] |
| Mar 4 | Zishen Wan | [TBD] | [TBD] |