New framework enables on-device AI training with float16 on RISC-V

Researchers have developed an open-source framework enabling on-device training of AI models using float16 precision on resource-constrained RISC-V single-core processors. This approach, which leverages standard RISC-V extensions like Zfh and Zvfh, can reduce memory footprint by approximately 50% with minimal impact on model performance. The framework also includes capabilities for layer-freezing to support transfer learning and fine-tuning scenarios, building upon the AIfES framework for embedded systems. AI

IMPACT Enables more efficient on-device AI model training, particularly for edge computing applications with limited resources.

RANK_REASON The item is an academic paper detailing a new hardware-software co-design for on-device AI training. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• float32
• Pierre-Alain Moëllic
• RISC-V
• Zvfh

• infra
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →