Can I read Zhang - DNN Training Under Variable Precision BFP With Stochastic Rounding on EtoBox?

Zhang - DNN Training Under Variable Precision BFP With Stochastic Rounding by Алексей Артамонов is a document available to read on EtoBox.

What is Zhang - DNN Training Under Variable Precision BFP With Stochastic Rounding about?

The paper presents the FAST system for DNN training using variable precision Block Floating Point (BFP) with stochastic rounding, which allows for efficient quantization and reduced training time. By incrementally increasing BFP precision across layers and iterations, the system achieves a 2-6× speedup in training while maintaining validation accuracy. The proposed architecture includes a FAST multiplier-accumulator (fMAC) that supports variable-width mantissas, optimizing hardware resource usage during DNN

Author

Алексей Артамонов

Language

EN