Cognitive SSD: A Deep Learning Engine for In-Storage Data Retrieval
summary
Cognitive SSD places a flash‑side deep learning and graph‑search accelerator called DLG-x inside the SSD, exposes functions via NVMe command extension, and reads model parameters and graph layouts directly from NAND to bypass DRAM/cache. An FPGA prototype with that accelerator and a NAND controller shows image‑retrieval latency down 69.9% latency reduction versus CPU on conventional SSDs and system power down up to 63.0% versus GPU solutions while keeping comparable accuracy.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:31Outline - Cognitive SSD
- 2:39Showcase Content Based Unstructured Data Retrieval System
- 3:26Solutions
- 4:06Cognitive SSD System--Overview
- 6:08Cognitive SSD System - Firmware and hardware
- 6:53DLG-x Accelerator - Deep learning unit
- 8:14DLG-X Accelerator-Graph search unit
- 9:57Cognitive System - Case study
- 11:41Evaluation Setup
- 12:20Evaluation-DLG algorithm
- 13:28Evaluation-Cognitive SSD System
- 13:58Evaluation-Cognitive SSD Cluster
- 14:59Conclusion
related talks
13mHardware/Software Co-Programmable Framework for Computational SSDs to Accelerate Deep...
14mGLIST: Towards In-Storage Graph Learning
24mManipulating Data in an Artificial Intelligence World
12mFlashNeuron: SSD-Enabled Large-Batch Training of Very Deep Neural Networks
01h 11mMachine Learning Acceleration in an Embedded World
22mCXL-ANNS: Software-Hardware Collaborative Memory Disaggregation and Computation...
19mSolidAttention: Low-Latency SSD-based Serving on Memory-Constrained PCs
12mBehemoth: A Flash-centric Training Accelerator for Extreme-scale DNNs