Pervasive and Sustainable AI with Adaptive Computing
summary
Uses RISC‑V microplatforms and FPGA-style adaptive fabrics with two open source toolchains called BAS and FIN to push DNNs into tiny embedded targets. Demonstrates 4→2→1‑bit quantization and pruning that cuts compute and memory by orders of magnitude (examples include a 666M inferences/sec kernel with 3 ns latency on 30 blocks and reported multi‑hundred to thousand× logic/power reduction). Targets low‑power embedded and real‑time latency constraints.
AI-written summary · about
chapters (19)
- 0:00Adaptive AI computing intro
- 5:23Generations of AI
- 10:41Generative AI risks
- 13:52Embedded AI requirements
- 16:04AI energy consumption
- 19:24AI efficiency engineering
- 21:40AI as signal processing
- 26:44AI vs traditional processing
- 31:22CNN hardware implementation
- 34:17FPGA for AI acceleration
- 36:24GPU and AI engine options
- 37:51Compute-memory balance in AI
- 38:15Data flow for power efficiency
- 39:58Quantization optimization
- 42:07Pruning network techniques
- 43:55AI primitives hardware
- 46:46Open-source AI hardware tools
- 49:24Adaptive AI benefits summary
- 51:39Q&A on AI's future
related talks
14mEdge AI: Real Solutions for Real World Problems
35mAI innovation begins with trust
30mBeyond GPUs: Why AI Agents Need a Need New Hardware
21mAhead of the Plasticity Curve: AI, Biology and Technology
18mDriving Innovation with AI
04m4 real-world AI deployments at enterprise scale
44mArtificial Intelligence in Small Embedded Systems
13mBuilding High-Performance AI Literature Processing