Key Highlights
- The rise of generative AI has shifted data centers into AI factories, where power infrastructure is a primary consideration.
- A transition to 800 VDC power distribution and integrated energy storage is proposed as the solution for addressing the challenges posed by high-power demands in AI workloads.
- Short-duration and long-duration energy storage systems are necessary to manage the volatility of power consumption in synchronous AI training environments.
- NVIDIA is collaborating with industry partners to develop standards and accelerate adoption of 800 VDC architecture in next-generation AI infrastructure.
From Data Centers to AI Factories: A Paradigm Shift in Power Infrastructure
The advent of generative artificial intelligence (AI) has transformed traditional data centers into bustling AI factories. These facilities now demand a fundamental shift in their power infrastructure, where efficiency and scalability are no longer secondary considerations but central to the design.
Challenges Posed by High-Power Demands
Traditional data centers were once vast halls of servers with little regard for power consumption. However, as AI workloads have grown in complexity and scale, these facilities must now grapple with unprecedented power demands. The relentless pursuit of performance has led to a significant increase in GPU power density, resulting in racks that consume tens or even hundreds of kilowatts, potentially scaling up to megawatt levels.
The NVIDIA Hopper to Blackwell architecture transition exemplifies this shift.
While individual GPU TDPs increased by 75%, the expansion of NVLink domains resulted in a 3.4x increase in rack power density, leading to a remarkable 50x performance boost. However, this leap also meant that traditional low-voltage systems (like 54 VDC) became impractical for delivering such high currents without significant resistive losses and copper cabling.
The Solution: 800 VDC Power Distribution
To address the challenges of high-power distribution, a dual-pronged approach is proposed: transitioning to an 800 Volts direct current (VDC) power distribution system alongside integrated energy storage. This new architecture offers several advantages:
- Native 800 VDC end-to-end integration can generate and deliver power directly to compute racks, reducing conversion steps and improving efficiency.
- This high-voltage DC (HVDC) approach supports higher density GPU clusters, driving greater performance per GPU and enabling more GPUs in AI factories.
- 800 VDC reduces the current required for a given power level, cutting copper use by 43% compared to traditional AC systems. This not only lowers material costs but also eases cable management as racks scale toward megawatt levels.
The transition to an 800 VDC architecture is not uncharted territory. The electric vehicle and utility-scale solar industries have already embraced HVDC to improve efficiency and power density, creating a mature ecosystem of components that can be adapted for data centers.
Managing Volatility with Multi-Timescale Energy Storage
While 800 VDC solves the efficiency-at-scale problem, it doesn’t address the volatility challenge. AI workloads introduce massive and rapid load swings due to synchronized GPU operations during training sessions. To mitigate these fluctuations, multi-timescale energy storage is essential:
- Short-duration storage: High-power capacitors and supercapacitors near compute racks can absorb high-frequency power spikes and fill valleys created by workload idle periods.
- Long-duration storage: Large battery energy storage systems (BESS) located at the utility interconnection manage slower, larger-scale power shifts.
These systems provide ride-through capability during transfers to backup generators.
The 800 VDC architecture is a key enabler for this strategy. By transitioning from traditional AC distribution to an 800 VDC model, it becomes easier to integrate storage in the most appropriate location within the data hall.
A Call for Collaboration: The Path Forward
This transformation cannot be achieved in isolation. Industry-wide collaboration is essential. Organizations like the Open Compute Project (OCP) provide a vital forum for developing open standards that ensure interoperability, accelerate innovation, and reduce costs for the entire ecosystem.
NVIDIA is collaborating with key industry partners across the data center electrical ecosystem to develop common voltage ranges, connector interfaces, and safety practices for 800 VDC environments.
The company has published a technical whitepaper on the 800 VDC architecture and plans to present details at the 2025 OCP Global Summit.
Companies interested in supporting this initiative can contact NVIDIA for more information. Together, these stakeholders aim to revolutionize AI infrastructure by making it more efficient, scalable, and capable of managing the power demands of modern AI workloads.
