面向下一代人工智能基础设施的800伏直流架构

Jared Huntington&Mike Tu The Architectural Imperative of 800 VDC and Integrated Energy Storage Contents Introduction...................................................................................................................................................3Increasing Power Demand for Higher Efficiency........................................................................................3Load Swings and Energy Storage...............................................................................................................5Grid Interconnect Requirements...............................................................................................................7Power Distribution Options.......................................................................................................................9800 VDC MGX Architecture......................................................................................................................13Datacenter Reference design...................................................................................................................16Industry Collaboration and Path Forward................................................................................................17Conclusion....................................................................................................................................................18 Introduction Only a few years ago, data centers were built aroundthecompute space, vast data halls of servers, withpower and cooling systems taking up a smaller share of the footprint. Then came the GPU revolution,transforming the data center into “AIFactories”.As showninFigure1,GPU racksare approaching100times greater power density compared to web serversand are increasing in powerat a near-exponentialpace,flippingthe balance.Power infrastructure, once secondary, now rivals or even exceeds the spacededicated to compute. As CPUs or GPUs are improved, there is typically an incremental increase in the GPU thermal designpoint (TDP) for power of 20% generation over generation. This leads to an increase in power requiredper server over time. Nvidia’s NVLink allows multiple GPUs to be networked together to effectively act asone large synchronous GPU and provide a significant improvement in performance when compared tooperating over Ethernet connections. This networking of GPUs is most effective when connected overcopper fromboth a power and cost standpoint, but at the expense of a limited reach due to signalintegrity. Since the highest performance can be achieved when more GPUs are on the same copperdomain with a limited reach, the maximum performance is tied to maximum power density. This meanspower increases are no longer 20% generation over generation, but can easily be 2x, 4x or 8x with theincrease in NVLink networking domain size. An exampleinFigure2aboveis the performance increase from Hopper to GB300. The TDP powerincreased by 75%, but the performance increased 50x with these changes. This also led to a 3.4xincrease in rack power density going from an 4x8 GPU NVLink domains (32 in the rack) to a 72 GPUNVLink domain. As GPU packing and packaging improves and networking topologies move to largerdomain sizes, this power density can continue to increase. This increase in power generation over generation and with NVLink domain drives a much more rapidincrease in power than has been seen in the past with GPUs. A secondary goal is to remove as manypower components from the NVLink domain radius as possiblesince this is the highest value real estatein the rack for performance. This combination of driving to higher power level and pushing the powercomponents away from the GPUs drives requirements for a different rack power architecture. To meet these unprecedented demands, 800 VDC has emerged as the optimal architecture for next-generation power distribution.It allows for minimizing theconversionandroutingvolume in thecomputespace while minimizingdatacenter distribution lossesand total end to end conversion stages.800 VDC significantly reduces current, copper use, and cable bulk compared to 54 VDC inside the rack orfacility level 480 VAC systems, while remaining safe and scalable. It benefits from the growing maturity ofSiliconCarbide (SiC)and Gallium Nitride (GaN)power conversion devices and the widespread adoptionof 800 VDC systems in theelectric vehicle (EV)industry.This enables seamless, end-to-end integrationfrom grid to rack, higher power density beyond 1 MW. Datacenters historically consisted of thousands of servers running diverse workloads. With GPUs theseworkloads can be synchronous across the datacenters leading to massive load swings in very short timeintervals. Energy storage can be used to mitigate these swings. This rapid shift in datacenter power infrastructure requirements puts the industry at a crossroads. To keep up with future GPU demands, weneed torethink how we deliver power. This leads us to the