AI Under the wave of computing power, lithium battery UPS How to refactor the power supply architecture of data centers
AI Intelligent Computing Center (AIDC) The power consumption of a single cabinet is higher than that of traditional cabinets IDC 4~8kW surge to 40~120kW, GPU Cluster hasPulsed, Severe fluctuation loadfeature; Tradition「Centralized power frequency UPS + lead-acid battery」The architecture occupies a huge amount of land, Insufficient response, Poor magnification performance, Can only be used as a passive backup, Unable to adapt to shortcomings such as elastic computing power. High rate lithium iron phosphate battery with a new generation of bidirectional modularization UPS, It's not just a simple battery replacement, But to promote the four major changes in the power supply system: Layered backup architecture, Integration of AC and DC routes, UPS Upgrade from emergency power supply to global energy storage node, The power supply system is transitioning from rigid static to flexible elastic, Thoroughly reconstruct from the power grid to GPU The complete power chain of the chip. Doledly Dongchi is a brand under Guangdong Beineng New Energy Co. , Ltd, Guangdong Beineng New Energy Co. , Ltd. was established in 2021 year, It is a company that integrates research and development, Production, sales, A professional manufacturer and service provider specializing in integrated data center services business, It is one of the world's most powerful integrated computer room equipment manufacturers .

One, Traditional data center power supply architecture AI Fundamental bottleneck in computing power scenarios
Classic architecture: 10kV mains electricity→transformer→400V Centralized power frequency communication UPS→Lead-acid battery pack→Low-voltage distribution cabinet→server power supply.
-
1, Mismatch of load characteristics
AI Training task load in a short period of time 30%→150% Drastic jump; Lead acid batteries have a common discharge rate≤0. 5C, Insufficient instantaneous high-power output capability, Extremely prone to voltage drop, GPU downclocking, Training task interrupted.
-
2, The spatial contradiction is irreconcilable
Lead acid has low energy density, The battery room occupies a large area of the computer room; In high-density intelligent computing scenarios, Each square meter of computer room is prioritized for deploying computing power servers, The huge opportunity cost between batteries is highlighted.
-
3, Poor lifecycle economy
Lead acid cycle life is only 300~500 time, 3~5 The entire group must be replaced annually; Data center design lifespan 10~15 year, at least 2 Next large-scale battery replacement, Overlay operation and maintenance, Temperature control cost, TCO Continuously rising.
-
4, single-function, only "Power outage safety net"
Traditional one-way UPS Only responsible for fault backup; Unable to participate in smoothing power grid fluctuations, Peak and valley arbitrage, Consume photovoltaic green electricity, Long term idle energy storage assets.
-
5, Rigid expansion mode
centralized UPS It must be constructed at once according to the long-term peak value; AI Staged production of computing power, Tidal fluctuations in computing power, A large number of equipment were unloaded in the early stage, Low asset utilization rate.
two, lithium battery UPS Bring about a leap in underlying capabilities: The core differentiation advantages of lithium batteries compared to lead-acid batteries (Intelligent computing scenario)
Mainstream solution: High rate lithium iron phosphate (LFP)
-
Power characteristics: Support 2C~6C Short term high rate discharge, perfect match GPU Instantaneous power peak; Partial cabinet level BBU Scheme adopted 10C The above ultra-high rate battery cells.
-
space: The energy density is about lead-acid 3 times, Reduce land occupation with the same backup power duration 40%~60%, Can achieve battery rack installation, Deploy nearby.
- lifespan: Cycle 3000~5000 time, design life 8~12 year, Match the lifecycle of data centers, Significantly reduce the frequency of replacement.
- System Capability: Supporting Intelligence BMS, Real time monitoring of individual battery cells, Fault Warning; Support frequent charging and discharging, Adaptive bidirectional inverter scheduling.
- Temperature adaptability: Wide temperature range operation, Reduce the continuous energy consumption of battery room air conditioning, Beneficial for the integrated deployment of liquid cooling equipment rooms.
shortcoming: Initial CAPEX higher; Safety control of lithium electric heating out of control, The fire protection plan requires significantly higher requirements than lead-acid.
three, Refactoring direction one: Power supply architecture - Centralized backup power from a single layer → Three level hierarchical distributed lithium battery power supply system
This is the most important architectural change, The industry is gradually taking shapePark level lithium batteries UPS + Computer room level lithium battery module + Cabinet level BBU (Battery backup unit) Three level collaborative protection architecture, Layered undertaking of power supply risks at different time scales.
hierarchy 1: park / Floor level|Megawatt level modular lithium battery UPS (second~Minute level backup power)
Positioning: Main and backup electrical barriers, Support diesel generator set start-up window period after mains power interruption (5~15min) .
-
Eliminate large power frequency machines, Fully adoptedSiC High frequency modular bidirectional lithium battery UPS; N+X Hot swappable redundancy, Staged construction of computing power, On demand expansion;
- Cancel the independent battery room, Lithium battery cabinet and UPS Placed side by side, Shorten the DC cable, Reduce line losses;
- Bidirectional topology: When the mains power is normal, it can be used as an energy storage system to participate in grid interaction, No longer just waiting for power outages.
hierarchy 2: Middle layer of computer room|Distributed lithium battery energy storage unit (millisecond~Second level fluctuation stabilization)
Regarding power grid flashover, Short term voltage disturbance, Compensate for concentration UPS Switching gaps, Smooth the overall power impact of the cluster, Reduce the upstream transformer and UPS Peak Capacity Configuration.
hierarchy 3: End of cabinet|Rack mounted lithium battery BBU (Microsecond level local support)
Directly deployed on the side of the server cabinet, distance GPU Recently.
effect: cope withMillisecond level instantaneous drop, Load pulse impact; When the superior power supply is temporarily abnormal, Local lithium batteries are immediately supported, prevent GPU Instantaneous power failure, Large model training breakpoint.
Industry trends: centralized UPS Responsible for long-term backup, End lithium battery BBU Responsible for instantaneous voltage stabilization and buffering, Collaboration between the two, No need to amplify upstream distribution capacity according to extreme peak loads anymore, Significantly reduce transformer, Cable investment.
Traditional architecture: All pressure is concentrated in one set UPS, One loss, all losses;
New architecture for lithium-ion batteries: Risk stratification and isolation, The impact range of the fault converges to the cabinet / cluster, Significant improvement in system resilience.
four, Refactoring Direction 2: Electric energy link - Promote communication UPS and 800V HVDC Integration of two high-voltage and direct current routes
AI Computing power has given rise to two mainstream power supply routes, lithium battery UPS It is an energy storage carrier for two routes:
route A: Modular communication of lithium batteries UPS (Stock renovation, The preferred choice for hybrid computing power parks)
10kV→400V Modular lithium battery for communication UPS→Communication and distribution→server power supply
- Advantage: Ecosystem maturity, Device Compatibility, Adequate operation and maintenance personnel; Suitable for renovating old and new computer rooms, General computing power cluster;
- Upgrade key: Host lithium battery adaptation + Bidirectional power conversion + ECO Efficient operation mode, The overall efficiency has been improved to 97%~98%.
route B: 800V HVDC HVDC Lithium battery system (New ultra large scale 10000 card intelligent computing cluster)
10kV→Rectification formation 800V DC bus→Lithium battery energy storage parallel DC bus→Server DC power supply
NVIDIA OCP White paper main promotion plan, It is also a large-scale overseas operation AIDC Mainstream Evolution Direction:
- reduce AC/DC Multiple transformation losses; Significant reduction in current of equal power, Save copper materials, Reduce line heating;
-
Direct parallel connection of lithium batteries to DC bus, omit UPS Inverter link, Simplified architecture, Reliability improvement; Core Conclusion: Regardless of communication UPS Architecture or high-voltage direct current architecture, All energy storage carriers will be electrified with lithium; Lithium batteries are no longer attached UPS host, Gradually becoming independent, Standardized energy storage units that can be flexibly connected to AC/DC busbars.
five, Refactoring direction three: Functional positioning transformation: UPS from「Passive emergency power supply」upgrade to「Computing Power Center Energy Dispatch Hub」
Tradition UPS: Normally standby, Work only after power outage, The idle rate of energy storage assets is extremely high.
lithium battery + bidirectional UPS expedite "UPS + Integrated energy storage" new model, A set of hardware that combines three major functions:
- Basic functions: Uninterrupted backup power supply (Original Mission)
-
Power grid support function: Load peak shaving and valley filling
AI There is a computing power tide in the cluster: Full load daytime training, No load at night. lithium battery UPS Charging during low electricity prices, Peak release of computing power, Reduce the maximum demand of the park, Save basic electricity costs.
-
Green electricity synergy function: Stabilize photovoltaic volatility
When supporting photovoltaics in the intelligent computing park, Lithium battery energy storage for intermittent green electricity consumption, Reduce dependence on municipal power grid, Empower PUE, Achieving carbon neutrality targets.
Fundamental changes: Energy storage is a cost item, Transforming into generating electricity revenue, Optimizing capacity allocation of operational assets.
six, Refactoring direction four: Reconstruction of Construction Mode and Operation and Maintenance System
-
deployment mode: prefabrication, modularity, Container type lithium battery UPS
The traditional on-site construction cycle is several months; Lithium battery modular system factory pre integration (UPS + lithium battery + firefighting + BMS Integrated pry block) , Quick on-site assembly, match AI Request for rapid project launch.
-
Capacity design logic innovation
Traditional design: According to the maximum peak value of a single cabinet × Total number of cabinets configuration UPS Capacity;
New Paradigm: Relying on instantaneous power support from lithium batteries, Design the main circuit according to the average load, Lithium battery hedge instantaneous peak value, Reduce the transformer, UPS installed capacity, Implement "Capacity reduction construction" .
-
Intelligent global collaborative operation and maintenance
UPS Monitoring System + BMS Battery Management System + Connect the computing power scheduling platform:
AI Algorithm predicts battery degradation, Risk of thermal runaway; Dynamically adjust battery charging and discharging strategies based on computing task scheduling; Implement predictive maintenance, Replace manual regular inspections.
seven, The core challenges facing the current landing
-
safety risk: Prevention and control of lithium electric thermal runaway in high-density computer rooms
The computer room is sealed, Equipment intensive, The risk of lithium battery fire and diffusion is higher than in ordinary scenarios; Need to be matched with liquid cooling temperature control, Specialized lithium battery fire protection, Cabin fire separation, Raise initial investment.
-
The standard system has not been fully unified yet
800V HVDC, rack-mounted BBU, Bidirectional lithium battery UPS Domestic and international standards are still iterating, Difficulty in interconnecting devices from different manufacturers.
-
Initial investment pressure
The procurement cost of lithium batteries is higher than that of lead-acid batteries; Higher short-term capital investment, The project must be based on5~10 Annual full lifecycle TCO modelPerform calculations.
-
Controversy over the Application of Tiered Batteries
Industry discussion on retired power batteries for new energy vehicles UPS Backup power, But consistency, safety responsibility, Overseas Certification (UL, EU Battery Regulations) Restrict large-scale promotion.
eighty percent, Prediction of Medium - and Long Term Industry Evolution Pattern
-
short-term (1~3 year)
New large and medium-sized construction AIDC lithium battery UPS Penetration rate continues to rise; Modular lithium battery for communication UPS and 800V HVDC Parallel development; "Concentrated lithium battery UPS + end BBU" Layered power supply scheme has become a mainstream design template.
-
mid-term (3~5 year)
UPS Continuous blurring of boundaries with energy storage inverters; Guangshu Zhirou has landed on a large scale in the intelligent computing park; Deep lithium battery energy storageParticipate in virtual power plants, demand response; Sodium ion batteries form a supplement in low to medium power backup scenarios.
-
forward
solid-state battery, supercapacitor + Lithium hybrid energy storage is gradually being piloted; Further simplification of power supply architecture, Ultimately forming a power grid — DC bus — distributed energy storage — Minimalist power link with computing power load.
AI The load characteristic revolution brought by computing power, Forcing the paradigm shift of data center power supply architecture. lithium battery UPS Not just upgrading battery materials, It achieves three major reconstructions:
-
Spatial restructuring: Energy storage moves from independent battery rooms to distributed and nearby deployment, Release computing power space;
-
Functional Refactoring: Upgrade backup power supply to dispatchable energy storage assets, Realize collaborative computing and electronics;
-
Architecture Refactoring: Single centralized protection moving towards multi-level distributed resilient power supply system.
Future oriented AI Design of power infrastructure for clusters, All must be done withHigh rate lithium battery energy storageAs a core variable, Re deduce the distribution capacity, bus voltage, Redundancy Strategy and Investment Return Model.