Beyond GPUs: Why AI Agents Need a Need New Hardware
summary
Argues that GPUs hit an efficiency wall for agentic AI and shows a different chip architecture: a reconfigurable data‑flow unit with a three‑tier memory on the SN40L RDU that keeps matrix ops on‑chip to cut back‑and‑forth and latency. Claims inferencing racks drop from 120–600 kW to about 10 kW and cites a 4x energy efficiency vs Blackwell-style systems; demonstrates a working agent pipeline on chat.sova.ai using large-context chains (tens of thousands of tokens) and faster end‑to‑end execution.
AI-written summary · about
chapters (7)
- 0:00Introduction and context
- 3:22Efficiency over raw intelligence
- 6:12GPU limitations and RDU architecture
- 12:28Energy efficiency metrics
- 18:03AI agents and their impact
- 24:51Live agent demo walkthrough
- 29:00Q&A and closing
related talks
47mBeyond Agents: What AI Strategy Really Needs in 2025
35mThe Fight for the Future of Artificial Intelligence
21mAnalyst Notebook
45mThe Curious Incident of Developing Artificial General Intelligence
21mAhead of the Plasticity Curve: AI, Biology and Technology
44mBuilding Agentic AI Systems: From Transactions To Conversations
10mKeynote Analysis
31mMyth & Truth behind Artificial General Intelligence