Get your free personalized podcast brief

We scan new podcasts and send you the top 5 insights daily.

Tyler Osborne describes how direct wafer-to-wafer bonding yields poorly when matching different functional dies because defects on either wafer randomly pair good silicon with dead components. Die-to-wafer bonding allows Intel to test each die beforehand and attach only known-good dies to functioning base positions.

Related Insights

As Moore's Law slows, the path forward isn't just smaller silicon transistors. Tan is investing in new materials like gallium nitride, silicon carbide, glass substrates, and even artificial diamonds to solve bottlenecks in advanced packaging and insulation, fundamentally changing chip architecture.

Kevin Fisher notes that Intel uses prioritized hotboxes to speed crucial problem-solving wafer lots through processing steps. However, assigning top priority to a single experimental lot forces production tools to idle or run inefficiently, behaving like a closed carpool lane that slows down normal wafer throughput.

Cerebras overcame the key obstacle to wafer-scale computing—chip defects—by adopting a strategy from memory design. Instead of aiming for a perfect wafer, they built a massive array of identical compute cores with built-in redundancy, allowing them to simply route around any flaws that occur during manufacturing.

Chris Oth explains that Intel's 10nm node slipped almost four years because engineers attempted hyperscaling without mature tools like EUV. Trying to innovate too much simultaneously caused massive schedule delays. Intel now enforces phase gates that defer unready breakthroughs to future generations to ensure customer reliability.

NVIDIA's approach requires connecting thousands of Grok chips, creating latency bottlenecks. Cerebras's CEO argues its single, integrated wafer-scale system avoids this "interconnect tax," offering superior memory bandwidth and performance for massive models by eliminating the wiring between thousands of tiny chips.

Intel's CEO emphasizes that the foundry business is fundamentally a service based on trust. While advanced nodes are crucial, customers will only commit if they have absolute faith in the foundry's yield, defect density, and cycle time, as a manufacturing failure is catastrophic to their own revenue.

Pushkar Arande points out that 80 percent of a modern semiconductor fab's cost structure is entirely fixed. Because of this massive overhead, small percentage improvements in machine utilization and wafer yields yield nonlinear revenue boosts, explaining why rising capital expenses have consolidated leading-edge manufacturing down to three foundries.

Unlike standard DRAM where products are standardized, HBM is less of a commodity. The complexity of manufacturing HBM—stacking multiple dice and advanced packaging—allows suppliers to differentiate on technology, yield, and thermal performance, giving them a competitive edge beyond just price.

The common thesis for Intel focuses on its process node recovery (18A, 14A). However, the critical bottleneck and new frontier for performance is advanced packaging—the ability to combine multiple silicon dies. This capability is the new driver of performance, effectively replacing the traditional Moore's Law of transistor shrinking.

Increasing the vertical dimensionality of chip packages heightens overall package value while making per-layer defect tolerances exponentially tighter. Because manufacturing flawless high-count stacks is mathematically and physically prohibitive, practical 3D implementations will likely remain capped at two to four layers.