Active Silicon Bridge Chiplets for Scalable XPU-Memory Integration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The limitations of lithography technology hinder the integration of high-performance computing components like processing units and high-bandwidth memory stacks, limiting the ability to increase compute power in silicon without compromising performance, and complicate the upgrade path for evolving AI workloads.
Innovation Solution
The use of active silicon-based bridge chiplets that integrate processing units and memory stacks, allowing for a modular and scalable assembly of AI bridge chiplets, which can be tailored for specific tasks and reduce processing burden on individual components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional silicon interposer-based integration is used, then manufacturing precision and reliability are maintained, but device complexity and manufacturing costs increase significantly
Solution Approach 1:
The system is divided into discrete bridge chiplet modules, each containing specific functionality (processor die, memory die, or interconnect logic). These segmented chiplets can be manufactured separately using standard CMOS processes and then assembled into larger systems, reducing the complexity of monolithic integration while maintaining manufacturing precision.
Solution Approach 2:
Bridge chiplets serve as intermediary components that mediate between processor dies and memory dies. These chiplets contain standardized interconnect interfaces that simplify the coupling between different functional blocks, reducing overall system complexity while enabling flexible configuration and lowering manufacturing costs through modular assembly.
2Productivity
If lithography technology is pushed to higher integration densities, then compute power increases, but manufacturing precision and reliability deteriorate due to process limitations
Solution Approach 1:
Instead of integrating all functions onto a single large chip that pushes lithography limits, the system segments functionality across multiple chiplets. Each chiplet can be manufactured at optimal density using current lithography processes without exceeding process capabilities, thereby maintaining manufacturing precision while achieving high overall compute power through parallel composition of multiple chiplets.
Solution Approach 2:
The patent transitions from two-dimensional planar integration to three-dimensional stacking architectures. Multiple chiplets are stacked vertically and interconnected through vertical vias and through-silicon vias (TSVs), enabling high-density integration without requiring excessive lateral lithography precision. This dimensional transition allows compute power to scale by adding layers rather than shrinking feature sizes.
3Adaptability or versatility
If AI bridge chiplets are customized for specific tasks, then adaptability improves, but device complexity increases
Solution Approach 1:
The bridge chiplet architecture employs universal standardized interfaces and common interconnect protocols that can be used across different application domains. While individual chiplets can be customized for specific AI tasks (e.g., neural network inference, training, or specific workload optimization), they all communicate through standardized interfaces, reducing configuration complexity and enabling reusable design IP across multiple product lines.
Solution Approach 2:
The system incorporates dynamic configuration capabilities where chiplet interconnections and resource allocation can be reconfigured at runtime based on workload requirements. This dynamic adaptability allows the same hardware platform to be optimized for different AI tasks without requiring complex static customization, balancing adaptability with manageable device complexity through software-controlled reconfiguration.
Data Source
AI summary
A first bridge apparatus comprising a connector circuit having a first interface to communicate with a processor die and a second interface to communicate with a stack of memory dies such that the processor die and the stack of memory dies are vertically adjacent to the connector circuit which bridges the processor die and the stack of memory dies. The first bridge apparatus further comprises a first connector link circuitry and a second connector link circuitry coupled to the connector circuit, wherein the connector circuit is a network-on-chip connector circuit. The first connector link circuitry is to communicate with a third connector link of a second bridge apparatus and the second connector link circuitry is to communicate with a fourth connector link of a third bridge apparatus.


