Multi-dimensional Memory Cluster with CXL Full-mesh Topology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory architectures, such as DDR, are not readily composable, making it challenging to construct large memory pools, which are necessary for advanced computing applications.
Innovation Solution
A multi-dimensional memory cluster system is developed with nodes connected in a full-mesh configuration, allowing for efficient allocation of memory based on service level needs, using Compute Express Link (CXL) connections and a novel addressing scheme that includes quality of service (QoS) and class of service (CoS) features, enabling deterministic latency control and fault isolation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If DDR memory with parallel interface is used, then memory capacity can be increased, but composability and ease of constructing large memory pools deteriorates
Solution Approach 1:
The memory system is segmented into multiple independent memory nodes (205) that can be individually constructed and then composed together. Each node contains memory devices and a controller, creating modular units that can be assembled into larger memory pools through standardized interconnects, thereby improving composability while maintaining scalability of memory capacity.
Solution Approach 2:
The memory nodes are designed with universal interfaces (CXL, PCIe) that enable them to function both as standalone memory units and as components of larger memory pools. The standardized connection protocols allow the same node design to be used in various configurations, from single nodes to large distributed memory systems, enhancing both composability and versatility.
2Speed
If memory nodes are connected in full-mesh configuration, then access speed and latency control improve, but system complexity increases
Solution Approach 1:
The system transitions from traditional two-dimensional memory组织 to an N-dimensional topology where nodes are arranged across multiple dimensions (e.g., 4 dimensions with 4 nodes each). This dimensional expansion allows full-mesh connectivity within each dimension while reducing overall system complexity through structured organization, enabling efficient access paths without requiring complete interconnection of all nodes.
Solution Approach 2:
The memory cluster is organized as nested structures where basic building blocks (4-node clusters) are nested within larger dimensional frameworks. Each nested level maintains full-mesh connectivity locally while interfacing with higher levels through standardized ports, achieving fast access within clusters while managing complexity through hierarchical organization.
3Adaptability or versatility
If service level descriptors are implemented for different nodes, then memory allocation adaptability improves, but information management complexity increases
Solution Approach 1:
Service level descriptors are pre-configured and stored in each memory node before runtime operations. These descriptors contain predetermined service level characteristics (fast, slow, etc.) that are discovered and cached by the host system during initialization, eliminating the need for complex real-time information management during memory allocation operations.
Solution Approach 2:
Each memory node autonomously maintains its own service level descriptor information, eliminating the need for a centralized information management system. The descriptors are self-contained within nodes and can be independently queried and updated, reducing overall system information management complexity while enabling adaptive memory allocation based on node-specific characteristics.
Data Source
AI summary
A multi-dimensional memory cluster. In some embodiments, the memory cluster includes a first node, having an external port for making a connection to a host; a second node, connected to the first node by a first memory-centric connection; the second node storing a service level descriptor; the first node being configured to: receive, from the host, a first request, addressed to the second node, for the service level descriptor; and forward the first request to the second node, the second node being configured to: receive the first request; and send a first response, to the first node, the first response including the service level descriptor of the second node.