Former Intel CEO: Why This is the Best Time to Build Hardware

影片總結

Former Intel CEO Pat Gelsinger says AI chip design is speeding up, but power, memory, and manufacturing will decide who scales.

Former Intel CEO Pat Gelsinger recounts joining Intel at 18, helping design the 386 and leading the 486, which pioneered modern electronic design automation. He says AI can accelerate chip design, but bringing a product to usable rack scale still takes about nine months after three months of design, with fabrication, advanced packaging, memory bandwidth, and power as major bottlenecks. Gelsinger calls HBM a poor but currently best-available memory and expects new memory technologies after three decades with no major breakthroughs. He predicts today’s roughly 100 AI accelerator vendors will consolidate as workloads, capital demands, and software favor a smaller set of platforms. He also warns that energy capacity limits data-center growth, predicting more project defaults, and argues that future virtualization must manage agents’ security, performance, migration, and policies.

章節

1. 0:00Intro: AI data centers face power limits as hardware becomes rack-scale
2. 1:00From tech school to Intel at 18: Gelsinger joined Intel at 18 and advanced to 486 design manager
3. 5:44The 486 and the birth of modern EDA: Intel's custom HDL and automated design tools
4. 7:23How AI changes chip design: Jalapeno points to AI-led logic design, while analog remains difficult
5. 9:18Why silicon still takes nine months: fabrication, packaging, racks, memory, and power constrain deployment
6. 14:49Will 100 AI chips converge to a few? Gelsinger expects workload shifts, capital needs, and buyers to narrow the field
7. 22:08Why HBM is a hideous memory: AI demand revives memory investment, but shoreline bandwidth remains a limit
8. 25:15How tall can chips get? Stacked memory promises progress, but eight- and 16-layer yields are difficult
9. 28:58How tall can chips get?: eight-layer packages, cooling limits, and costly optical memory links
10. 34:07How tall can chips get?: optical IO, NVL72, and the 2028–2029 transition
11. 38:33How tall can chips get?: optical switching, converging networks, and AI workload trade-offs
12. 42:49Energy capacity equals economic capacity: US power limits, 800-volt DC, and new hardware
13. 48:42A VMware for agents: virtualized agent management, security, and human-set policies