Modular Compute Storage Clusters via Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional compute and storage infrastructure is inefficient in supporting rapidly changing workload requirements in AI and cloud computing, as it struggles with increasing computing performance, data storage capacity, and communication bandwidth, leading to inefficient capacity expansion and hardware utilization.
Innovation Solution
A modular architecture for compute and storage hardware that allows for easy configuration, repurposing, and capacity expansion, with uniform baseboards for different components like CPU, GPU, SSD, and ASIC, interconnected via high-speed networks, enabling dynamic reconfiguration and resource pooling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional server package method is used to support new workloads, then hardware system can be configured, but resource utilization becomes inefficient and capacity expansion is costly
Solution Approach 1:
The patent segments the server system into independent functional modules: compute modules (with CPUs, GPUs, FPGAs), storage modules (with SSDs, HDDs), and network modules. Each module can be independently configured, added, or removed from standardized racks, allowing efficient resource utilization by matching exactly the components needed for specific workload requirements without deploying entire traditional server units.
Solution Approach 2:
The patent creates universal standardized racks and module interfaces that can accommodate different types of compute, storage, and network modules. This universal platform supports multiple workload types (AI training, inference, cloud computing, storage) using the same infrastructure, improving adaptability while maintaining efficient resource utilization through consistent management and allocation.
2Power
If additional server units are added for capacity expansion, then computing capacity increases, but facility equipment requirements and costs increase proportionally
Solution Approach 1:
The patent merges multiple computing modules into shared standardized racks that consolidate facility equipment resources. Power supplies, cooling systems, and network infrastructure are shared across multiple compute, storage, and network modules within the same rack, so adding computing capacity does not require proportional increases in facility equipment - the infrastructure is pooled and reused.
3Power
If single chip performance is increased with more cores, then processing capability improves, but performance bottleneck is reached due to slowing Moore's Law
Solution Approach 1:
The patent segments processing capability across multiple independent modules rather than relying on single-chip performance. By distributing workloads across multiple CPUs, GPUs, and FPGAs in separate modules within standardized racks, the system achieves scalable processing capability that adapts to different workload environments without being constrained by single-chip performance bottlenecks.
Solution Approach 2:
The patent transitions from vertical scaling (increasing single-chip cores) to horizontal scaling (adding multiple modules with fewer cores each). This dimensional shift allows the system to achieve required processing capability through parallelism across multiple modules, bypassing the Moore's Law bottleneck while maintaining adaptability to various workload requirements.
4Productivity
If heterogeneous hardware systems are created for different workloads, then specific workload performance is optimized, but system complexity and platform diversity increases
Solution Approach 1:
The patent segments different workload optimizations into separate standardized modules rather than creating entirely different hardware systems. Each module type (CPU-based, GPU-based, FPGA-based, storage-optimized) is independently designed and tested, then integrated into the same standardized rack infrastructure, achieving workload-specific performance without proportional increases in overall system complexity.
Solution Approach 2:
The patent creates a universal standardized rack platform that can accommodate multiple heterogeneous module types. The same rack infrastructure, interfaces, and management system support AI training modules, AI inference modules, cloud computing modules, and storage modules, reducing platform diversity complexity while maintaining optimized performance for each workload type.
Data Source
AI summary
A computing and storage system includes a housing having power and cooling facility and a plurality of slots, each slot having interface connection. An interconnecting board is coupled to the interface connection of the plurality of slots. A plurality of baseboards are inserted, each in one of the slots wherein a board interface mates with the interface connection. All of the baseboards have the same form factor and the same board interface and each of the baseboards has a plurality of electronic devices, such that all of the electronic devices mounted onto one of the baseboards are the same. Using the interconnecting board, the various baseboards can be interconnected to form a computing and/or storage machine of different operational characteristics as required by a given task. In addition, the inter connection board is managed to adjust the networking resource allocations for different traffic characteristics and workload requirements.