1 min read
White Paper: GPU-Initiated, Liquid-Cooled, Ultra-High-Density Storage for Next-Gen AI
This paper introduces a paradigm shift in storage architecture designed to overcome the CPU-centric data path...
High Power Architecture
The Backbone of Next-Gen AI Clusters: Defining the Future of Resilient, Scalable, and High-Density Power Infrastructure.
AI Data Center Power: Engineering the System-Level Revolution
As Generative AI pushes compute demands beyond physical limits, server power architecture is undergoing a system-level revolution. As rack power density accelerates from 20 kW to beyond 100 kW—and with proposals reaching up to 1 MW, power architecture has become the critical backbone of AI computational performance.
Wiwynn’s Strategic Pillar: Core Technologies Defining Hyperscale Compute
Vertical Power Delivery (VPD)
As AI accelerator currents exceed 1,000A, traditional lateral power delivery hits a "Power Wall" of prohibitive voltage drops. Vertical Power Delivery (VPD) solves this by mounting conversion modules directly beneath the processor.
• Path Loss Mitigation: Ultra-short vertical paths slash PDN I^2R losses by >80% at high-current AI loads.
• Architectural Freedom: Backside delivery frees up the top-side PCB for additional HBM (High Bandwidth Memory) or reduced form-factor designs.
• Voltage Integrity: Ultra-low inductance minimizes voltage droop during extreme load transients, protecting signal integrity.
• Thermal Optimization: By minimizing resistance-induced heat, VPD enhances the overall effectiveness of liquid cooling loops.
Scalable Rack-Level Power
The transition to Modular Rack-Level Power is essential for AI elasticity. Our integrated architecture—featuring centralized busbars and intelligent power shelves—delivers the high-availability redundancy required for absolute stability under extreme transient AI workloads.
• Exponential Capacity Scaling: Seamlessly scaling from 190kW (V2) to 300kW (V3), and reaching a massive 1.1MW (V4) to meet the most demanding AI training clusters.
• Disaggregated Topology: Evolving from single-rack integration to disaggregated Power Racks and IT Racks (V3/V4), enabling modular deployment and superior data center floor-plan flexibility.
• Advanced HVDC Transmission: Transitioning to ±400VDC or 800V architectures to neutralize transmission losses and maximize end-to-end energy efficiency.
• Dynamic Transient Response: Engineered for the "bursty" nature of AI workloads, utilizing Capacitor Backup Units (CBU) to ensure seamless voltage stability during rapid compute load fluctuations.
Next-Gen Power Integration: SiC, GaN, and SST
We are redefining power delivery by leveraging SiC/GaN semiconductors and Digital Twin insights. This architecture achieves deep synergy with advanced liquid cooling solutions, creating a "smart" energy chain for the entire AI infrastructure.
• Centralized Power Nodes: The Power Rack acts as the intelligence hub, converting grid-scale AC to stable HVDC. Integrated BBU/CBU energy storage provides a robust backup buffer, reducing server-level complexity.
• Solid-State Transformer (SST): Our SST technology provides a high-density alternative to traditional bulky transformers. By converting power directly to HVDC at the facility level, SSTs offer superior control flexibility and modular scalability for rapid hyperscale deployment.
