The Circuit Episode 121: Stop Talking about SAM Please
32 snips
Jun 9, 2025 Ben and Jay dive into the booming semiconductor industry, spotlighting Broadcom's impressive earnings linked to the AI surge. They tackle the hurdles of ASIC development and the rising significance of innovative networking, particularly with the Tomahawk 6 chip. The chat shifts to the competitive landscape among hyperscalers, discussing the evolving dynamics in chip design and processing capabilities. Finally, they explore the essential role of storage solutions for AI, stressing the need for powerful infrastructures to manage increasing data demands.
AI Snips
Chapters
Transcript
Episode notes
Broadcom's AI ASIC Growth Potential
- Broadcom's AI ASIC revenue could reach $30 billion in 2027, driven mainly by three major customers.
- However, many ASIC projects fail to scale if they don't outperform NVIDIA's GPUs on cost and performance.
ASIC Design Challenges and Timeline
- Designing custom ASICs for AI takes about two years from start to tape-out and shipment.
- Customers must achieve a meaningful performance and cost improvement over GPUs for projects to continue scaling.
AI Chip Architectures Align
- Google's AI chips like Tranium and NVIDIA's Blackwell are architecturally similar, focusing on tensor core units over traditional GPUs.
- This shifts the AI compute focus from gaming GPU designs to matrix multiplication efficiency.