Hardware/Software Co-Programmable Framework for Computational SSDs to Accelerate Deep...
summary
HolisticGNN implements a FPGA-based computational SSD that stores graphs as neighbor-oriented adjacency lists, exposes RPC over PCIe and runs programmable GNN operators where data live. Hardware RTL and software run on an FPGA-CSSD prototype across 13 datasets; the system converts CSR to adjacency lists, separates high/low-degree vertices for compact updates, and executes batching and sampling near-storage. Measured end-to-end inference averages 7.1× faster than modern GPUs while cutting energy use by 33.2×.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:30Summary
- 1:14Main Talk
- 2:35Genome Preprocessing
- 3:20Genetic Processing
- 4:40Holistic GNN
- 6:15Graph Runner
- 8:50Evaluation
- 9:55Demo
related talks
14mGLIST: Towards In-Storage Graph Learning
16mCognitive SSD: A Deep Learning Engine for In-Storage Data Retrieval
09mAccelerating compute with software defined hardware (FPGA's)
23mSparse Computing Drives Energy-Efficient Computation for Artificial Intelligence
01h 52mNeural Network Accelerator Design
43mHigh Speed Query Execution with Accelerators and C++
01h 11mMachine Learning Acceleration in an Embedded World
22mCXL-ANNS: Software-Hardware Collaborative Memory Disaggregation and Computation...