AI custom Software/Hardware Interface improving performance 5-10x
summary
Customizing a RISC‑V core via Codasip Studio’s Codasip Fusion adds bounded micro‑accelerators (about 3,000 lines of RTL) to speed a sensor + MCU anomaly detector: an AI Mac instruction for 8‑bit MACs and a new single‑cycle unaligned loads op are created from a software model, about 40 custom ops in ~1 month, yielding ~4.36× application speedup (claims up to 5–10×) and 4.36× performance and 7× lower power measured post‑synthesis on RTL while preserving core verification.
AI-written summary · about
chapters (9)
- 0:00Introduction and overview
- 1:05AI at the edge project
- 2:12Performance challenges and goals
- 3:15Fusion software interface
- 4:59Custom instruction creation
- 6:22Auto-aligning instruction demo
- 7:28Performance improvements achieved
- 8:55Power reduction results
- 9:155-10x performance gains summary
related talks
12mCustomizing RISC-V Cores to Accelerate Neural Networks
13mDemo: RISC-V as an Enabler of Heterogeneous Compute
01h 10mWhy RISC-V Through ISA Customization Demonstration
19mEnhancing RISC-V Embedded Processor Performance Through Advanced Instruction F...
18mRISC-V Performance Delivered
19mAccelerating Edge AI with the Help of RISC-V
15mOptimizing Real-Time Application Requirements on ARC-V Processors Leveraging RISC-V Ex...
22mAdvancing AI Computing With Optimized Matrix Multiplication Techniques With RISC-V CPU