Onsemi Promotes 800 VDC Power Architecture for AI Infrastructure

Onsemi, a leading semiconductor company, is actively advocating for the adoption of an 800 VDC (direct current) power architecture as a standard for future artificial intelligence infrastructures. This initiative comes amidst a growing demand for power and efficiency in data centers hosting AI workloads, ranging from Large Language Models (LLM) to complex neural network training.

Evolving computational requirements, driven by advancements in AI, necessitate increasingly robust and performant power systems. Onsemi's 800 VDC approach aims to address these challenges, offering significant potential in terms of reducing energy losses and optimizing power density within server racks. For CTOs, DevOps leads, and infrastructure architects, the choice of power architecture represents a strategic decision with direct impacts on Total Cost of Ownership (TCO) and operational sustainability.

Technical Details: 800 VDC and Its Advantages

The 800 VDC architecture distinguishes itself from traditional alternating current (AC) power systems or lower DC voltages (such as 48 VDC). Adopting a higher direct current voltage allows for a reduction in the current required to deliver the same power, minimizing Joule losses along cables and within conversion components. This translates into greater overall energy efficiency for the data center.

Furthermore, an 800 VDC architecture can simplify the power conversion chain, reducing the number of stages and the complexity of Power Supply Units (PSUs) within servers and racks. This simplification not only improves reliability but also enables higher power density, which is essential for accommodating an increasing number of high-performance GPUs – such as NVIDIA H100s or AMD Instinct MI300X – in limited physical spaces. Thermal management, a critical aspect for AI systems, can indirectly benefit from reduced heat dissipation due to improved efficiency.

Implications for On-Premise AI Infrastructure

For organizations evaluating on-premise or self-hosted AI infrastructure deployments, the power architecture is a key factor. The ability to deliver power efficiently and densely directly impacts operational costs (OpEx) related to energy consumption and cooling. An 800 VDC system, by reducing losses, can help contain energy bills, an increasingly relevant aspect given the power demands of modern AI workloads.

Data sovereignty and regulatory compliance often push companies towards on-premise or air-gapped solutions. In these scenarios, complete control over the physical infrastructure, including power, becomes fundamental. Adopting standards like 800 VDC can facilitate the design of more compact and powerful data centers, allowing for the allocation of more computational resources (such as VRAM and compute cores) per unit of space. This is particularly advantageous for those managing large LLMs or intensive training pipelines with specific latency and throughput requirements, without relying on external cloud resources.

Future Outlook and Considerations for Decision Makers

Onsemi's initiative highlights a clear trend in the industry: optimizing physical infrastructure is as crucial as advancing AI algorithms and models. The adoption of innovative power architectures like 800 VDC will require collaboration among component suppliers, server manufacturers, and data center operators to ensure interoperability and standardization.

For technical decision-makers, evaluating these new technologies involves a thorough analysis of trade-offs. While 800 VDC promises greater efficiency and density, it may also require initial capital expenditures (CapEx) for upgrading existing infrastructure. It is essential to consider the entire lifecycle of AI deployment, from hardware provisioning to operational management. For those evaluating on-premise deployments, AI-RADAR offers analytical frameworks on /llm-onpremise to assess the trade-offs between different infrastructural solutions, helping to make informed decisions that balance performance, TCO, and data sovereignty requirements.