Building reliable AI hardware
summary
Semiconductor demand heads toward a trillion dollar market as AI moves to the edge and into vehicles, forcing custom silicon to solve reliability. Data centers running millions of GPUs see roughly one failure every 30 minutes at scale, so AWS-style in-system fleet monitoring that pinpoints failing transistors and a Siemens-like approach embedding sensors, telemetry (temperature, degradation, performance) and a secure element to encrypt diagnostic channels enable predictive failure to remove parts before they crash multi-week LLM training jobs.
AI-written summary · about
chapters (10)
- 0:00AI hardware reliability matters
- 3:41Data center reliability challenges
- 6:55AI in automotive safety
- 10:07Modern AI hardware chips
- 12:57Fleet monitoring for AI chips
- 15:27GPU failure probability
- 17:12Reducing failure blast radius
- 18:18Silicon degradation monitoring
- 19:22AI hardware diagnostics risks
- 22:15Secure AI hardware
related talks
47mEmbedding AI: Don’t Burn Your Fingers in the Heat of the AI revolution
06mManaging and Securing AI Output
29mTrustworthy AI for the Physical World
57mKeynote: Information Technology Systems in a Post-Silcon World
20mTrusted Hardware for AI Workloads: Extending Confidential...
09mThe new era of smart machinery-utilizing edge intelligence to create an intelligent world
44mGPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud
15mEnsuring 100% uptime in the cloud of failing GPUs