A network-centric hardware/algorithm co-design to accelerate distributed training of deep neural networks

Keyphrases

• Aggregator 100%
• Network-centric 100%
• Hardware-software Co-design 100%
• Deep Neural Network 100%
• Distributed Training 100%
• In-network 75%
• Training Algorithm 75%
• Neural Network 50%
• Process Training 50%
• Network Interface Card 50%
• In-network Computing 50%
• Aggregation Nodes 50%
• Training Information 50%
• Data Compression 25%
• Compress 25%
• Order of Accuracy 25%
• Multiple Streams 25%
• Lossy Compression Algorithms 25%
• Hardware Efficiency 25%
• Unique Value 25%
• Value Characteristics 25%
• Distributed Systems 25%
• Floating-point 25%
• Communication Cost 25%
• Associative Property 25%
• Communication Time 25%
• Training System 25%
• Conventional Training 25%
• Gradient Update 25%
• Gradient Communication 25%
• Aggregation Operators 25%
• Embedded Data 25%
• Precision Loss 25%
• Eon 25%

Computer Science

• Participatory Design 100%
• training algorithm 100%
• Deep Neural Network 100%
• Training Process 66%
• Neural Network 66%
• Network Interface Card 66%
• Distributed Systems 33%
• Data Compression 33%
• Floating Point 33%
• Communication Cost 33%
• Training System 33%
• Multiple Stream 33%
• Embedding Data 33%
• Direct Implementation 33%
• Lossy Compression Algorithm 33%
• Characteristic Value 33%