Faris Sbahi, Normal Computing | CASPA CEO Summit 2026
summary
Normal Computing claims HBM is the scaling bottleneck — it uses three to four times more wafers than DRAM and is effectively sold out through 2028 — and proposes a compute-in-memory rack strategy to slash joules per token. Roadmap names Kelvin (first rack system shipping in 2028) and Stirling (custom memory co-design) target roughly two orders of magnitude gains in tokens per second per watt through 3D stacking and tier‑1 memory partnerships including Samsung Catalyst.
AI-written summary · about
chapters (7)
- 0:00AI Semiconductor Innovation and Capacity Constraints
- 3:45Overcoming Memory Bottlenecks With In-Memory Computing
- 6:20Co-Designing Complete Inference Systems
- 9:40Physical AI and Edge Efficiency Challenges
- 11:59AI Product Roadmaps, Value, and Public Trust
- 15:58Cheap Intelligence as an Industrial Resource
- 18:37Scaling AI Through Chips and Customer Partnerships
