AI summary by TwinMind
Recapping the AI and Semiconductor Symposium
Asianometry
9m 39s
Event Origins & Planning
- Symposium held September 3rd, 2023 in Taipei
- Idea originated when Dylan and Doug O'Loughlin (Fabricated Knowledge) both visiting Taipei
- Built on success of previous May meetup between Dylan and host
- Venue: NTU Hospital International Convention Center ballroom
- First announcement made week after booking in early August
- Over 400 signups by sales close
- Nearly 200 people attended despite short notice and typhoon threat
Pre-Event Challenges
- Host experienced anxiety coordinating slides, seating, logistics
- Typhoon Hai kuei approaching Taiwan for several days
- Storm fortunately turned south, avoiding direct hit on Taipei
- Dylan sprained ankle drunk dancing week prior, walking on it whole week
- Emergency visit to traditional Chinese medicine clinic 1.5 hours before event
- Treatment took 20 minutes, causing additional stress
- Dylan arrived 20 minutes late to venue
- 30 people already waiting at 12:30 PM
- Rita (master of ceremony) arranged car ride from Beitou to convention center
Talk 1: Frankwell Lin (Andes Technology Group)
- Speaker: Chairman and CEO of Andes Technology Group
- Also serves on RISC-V International Association board
- Company: Semiconductor IP company providing IC design services
- Develops hardware and software products
- Topics covered:
- AI accelerators including photonics-based ones
- Strong position in RISC-V
- Methods for continued performance improvements in digital electronics
- Moore's Law challenges
- New materials and devices (time-consuming approach)
- Architectural specialization
- "Extreme heterogeneity" - very specialized chips for specific applications
- Universal Chiplet Interconnect Express (UCIe) chiplet standard
Talk 2: Dylan's AI Computational Requirements
- Metaphor: BERT model parameters printed on paper would stack from Philadelphia to Chicago
- GPT-4 parameters would stack to moon and back 10 times
- Topics:
- Model requirements for hardware and future scaling
- Future model architecture and hardware needs
- Memory hierarchy walkthrough
- Training costs for largest leading-edge models
- Cost comparison: ChatGPT vs Google search
- Audience behavior: People lifting phones to photograph each slide
Talk 3: Warren Lao (Alethea Capital)
- Speaker: Senior analyst from Hong Kong-based Alethea Capital
- Focus: AI boom and AI supply chain investment thesis
- Key points:
- ChatGPT and Microsoft Copilot require massive AI server installations
- High-end server cost: ~$180,000
- GPU captures 89% of server cost
- Nvidia prominently featured but many other suppliers involved
- Discussion of AMD's upcoming MI300 processor (first CPU-GPU combo)
- Comparison with H100 accelerator
- Insight into massive vendor network feeding ChatGPT
Talk 4: Doug O'Loughlin's China Lagging-Edge Thesis
- Topic: China's push into lagging-edge semiconductor industry
- Supporting thesis elements:
- Great wave of digitization in automotive industry requiring more silicon
- China holds substantial cost and market advantage in EV production
- Leveraging advantages to generate immense capacity for EV exports
- Many EV makers vertically integrated, building fabs for internal use
- Distorts whole system economics
- Chinese unable to import leading-edge semiconductor manufacturing equipment
- Complete shift toward dominating lagging-edge
- Implications: Traditional automotive semiconductor OEMs will face hard competition from new wave of Chinese capacity
- Audience reaction: Provocative thesis, several attendees visibly moved
Host's Talk: American Lithography Companies
- Story about American lithography companies
- Full version video to be released later
- Opening anecdote:
- First TSMC video performed better than previous content
- Discovered top search term was "TSMC Minecraft"
- Large YouTuber "TSMC Minecraft" makes Minecraft tutorials
- Significant portion of views from people seeking Minecraft content, not semiconductor company
Panel Discussion
- Started early due to Dylan's incomplete slides for second talk
- Jordan Schneider (ChinaTalk) participated despite jet lag
- Used Slido service for question collection and upvoting
- Unable to answer all received questions
- Notable moments:
- Dylan asked about possible AI server bust
- Rant about Jensen's tendency to double chip orders when going is good
- Prediction of eventual bust and inventory write-downs
- Dylan's "GPU poor" concept from prior newsletter
- Focus on Google's massive computing advantage
- References to Google's MEMS-based optical switches for dynamic traffic rerouting
- Former Google employee confirmed via chat: "everything he is saying about Google is pure Google thinking"
- Showed chart that prompted Sam Altman to call Dylan "Google marketing/recruiting" on Twitter/X
- Elon Musk replied to Altman's tweet
Event Conclusion
- Finished exactly on time at 5 PM despite late start
- Room available until 6 PM for mingling
- Audience gravitated toward Doug and Dylan (described as "legitimate celebrities")
- Memorable moment: Someone brought real wafer (big one) for three speakers to sign
- Dylan able to identify chips on wafer while signing
- Host suffered back spasms from event stress but enjoyed talking with attendees
- Nearly 200 attendees despite short notice and typhoon conditions
