Full Transcript
Loading transcript…
INTEL KEYNOTE: Advanced packaging as the critical enabler. WHAT ARE THE SUCCESS FACTORS?
Watch with subtitles, summary & AI chat
Add the free Subkun extension — works directly on YouTube.
- Watch
- Subtitles
- Summary
- Ask AI
Engineers, researchers, and industry professionals interested in the future of AI hardware and semiconductor packaging.
TL;DR
Intel's Nick Sila discusses the critical role of advanced packaging in enabling co-packaged optics (CPO) for AI systems. He highlights how CPO addresses the IO bandwidth bottleneck limiting AI compute, enabling larger scale-up domains and preparing the supply chain for massive demand.
Key Takeaways
- AI compute growth has outpaced memory and IO bandwidth, creating significant bottlenecks that limit system scalability and performance.
- The primary challenges for scaling AI systems are the bandwidth density and overall efficiency of the input/output interfaces.
- Optics, particularly co-packaged optics (CPO), are essential for enabling larger scale-up domains by overcoming the reach limitations of copper interconnects.
- The demand for optical links in next-generation data centers is projected to be massive, requiring significant supply chain preparation and investment.
- Advanced packaging techniques, like heterogeneous integration and 3D stacking of chiplets, are crucial for scaling compute performance beyond single-die limitations.
- CPO integration benefits from advanced packaging by offering lower power consumption, increased bandwidth, improved thermal management, and reduced signal path lengths.
- The industry needs to prepare for a substantial ramp-up in CPO adoption, impacting everything from component supply to assembly and metrology capabilities.
- Co-packaged optics represent a complementary technology to copper, facilitating a gradual transition towards a more optical future for entire systems.
In This Video
- 00:00Introduction to European Industry Success
The keynote begins by highlighting a European industry success story, Optoscribe, acquired by a semiconductor leader.
- 01:02Scaling Co-Packaged Optics Challenges
The speaker will discuss challenges in scaling co-packaged optics and industry efforts to solve them.
- 01:33AI Compute Bandwidth Bottlenecks
AI compute growth outpaces memory and IO bandwidth, creating 'walls' and limiting system scalability.
- 02:48Bandwidth Density and Efficiency Needs
Scaling AI systems requires improvements in bandwidth density and efficiency of IO, especially in scale-up networks.
- 03:54Optics Enable Larger Scale-Up Domains
Optics overcome copper limitations, enabling larger scale-up domains and supporting future CPO deployments.
- 05:04Staggering Demand for Optical Links
Next-generation data centers require millions of optical links, demanding significant supply chain readiness.
- 07:03Advanced Packaging for Scalability
Advanced packaging, using heterogeneous integration and chiplets, is crucial for scaling compute performance.
- 08:01CPO Integration in Advanced Packaging
Co-packaged optics must integrate into the advanced packaging ecosystem, offering benefits like lower power and better thermals.
Questions & Answers
What is the main challenge in scaling AI systems?
The key challenge is the bandwidth density and efficiency of the IO itself, as memory and IO bandwidth have not kept up with the growth in compute, leading to systems becoming IO limited.
Why is copper facing challenges in scaling network domains?
Copper is running into significant challenges at higher speeds, and while mitigations exist, they add power consumption and are still limited in reach, hindering the ability to scale up systems.
How does optics help in scaling network domains?
Optics enables scale-up domains to no longer be limited by copper's reach, allowing for larger domains and more links between XPUs, ultimately driving system performance.
What is the projected demand for optical links in data centers?
Andy Betostein projected more than 100 million 600G equivalents in a 1 gigawatt data center, with exponential growth expected for the rest of the decade.
How does advanced packaging help in scaling compute?
Advanced packaging enables scaling by moving to multi-die packages due to reticle limitations with single dies, facilitating heterogeneous integration of specialized chiplets.
What advantages does advanced packaging bring to CPO?
It enables lower power consumption, increased bandwidth, better thermal management through direct cold plate integration, and reduced signal path lengths due to 3D stacking.
Key Terms
- Co-packaged Optics (CPO) — A technology that integrates optical components directly within the same package as the electronic components, like CPUs or GPUs.
- Memory Wall — A bottleneck where the speed of memory access does not keep pace with the speed of processors, limiting overall system performance.
- IO Bandwidth — The rate at which data can be transferred to and from input/output devices or interfaces in a computer system.
- Chiplets — Smaller, specialized integrated circuits designed to be combined in a single package to create a larger, more powerful system.