AI 数据中心温控与安全储能Thermal Management and Safe Energy Storage for AI Data Centres
1. 海湾与园区1. The Gulf campus
海湾正午。白色模块连成一列,室外按五十度来理解这座园区要面对的热。Noon in the Gulf. White modules in a line. The campus is read against an outdoor temperature of fifty degrees.效果图_09_高温园区.jpg园区整体:数据大厅、屋面冷源、电力模块、围栏内储能,以及停车场光伏雨棚。一个模块二十兆瓦,可以复制到整座园区。The campus: data halls, roof cooling plant, power modules, fenced storage, and solar canopies over the car park. One module is twenty megawatts, and it can be repeated across the site.效果图_01_园区整体.jpg暮色里灯还亮着。热被带走,事故留在单柜,算力不停。Dusk, and the lights stay on. The heat leaves, the fault stays in one cabinet, and the compute keeps running.效果图_13_暮色园区.jpg
2. 数据大厅2. The data hall
数据大厅:高密机柜列、二次侧管路、列端 CDU,右侧风墙接住大约百分之十五的热。The data hall: high-density rows, secondary piping, CDUs at the end of the row, and a fan wall taking about fifteen percent of the heat.效果图_02_数据大厅液冷.jpg新一代 GPU 机柜,单柜大约一百二十到一百三十千瓦。大约百分之八十五的热从冷板进入液体。A new-generation GPU rack, about 120 to 130 kilowatts. About eighty-five percent of the heat enters the liquid through the cold plate.效果图_10_GPU机柜.jpg热离开机柜,走的是液体,不是风。The heat leaves the rack in the liquid, not in the air.效果图_15_液体离开.jpg列端是佳力图 CDU,单台一点五兆瓦级。一个模块十二用、二备,共十四台,N 加二。Canatal CDUs at the end of the row, 1.5 megawatts each. Twelve on duty and two on standby, fourteen in all, N plus 2.效果图_18_列端CDU.jpg风墙接住剩下的热,也给网络和存储散热。十七台,十五用、二备,和 CDU、屋面冷水一起群控。The fan wall takes the remaining heat, and cools the network and the storage. Seventeen units, fifteen on duty and two on standby, controlled with the CDUs and the roof chilled water.效果图_14_风墙.jpg
3. 冷板与二次侧3. Cold plate and secondary loop
冷板贴在 GPU 上。不锈钢歧管和自密封软管可以热插拔。The cold plate sits on the GPU. Stainless manifolds and self-sealing hoses can be swapped while the rack is live.效果图_05_冷板与二次侧.jpg柜底有接液盘,也有漏液监测。A drip tray and leak detection sit at the bottom of the rack.效果图_20_接液盘.jpg不锈钢二次侧把冷却液送到每一柜。一个二十兆瓦模块,这条路大约覆盖一百五十四柜。Stainless secondary piping carries coolant to every rack. On one 20-megawatt module, the path covers about 154 racks.效果图_19_二次侧管路.jpg事故要留在这一柜里,数据大厅不用停。An incident has to stay in this one cabinet, so the data hall does not stop.效果图_16_单柜.jpg
4. 屋面冷源4. Roof cooling plant
屋面冷水机组提供一次侧冷冻水,在 CDU 里和二次侧冷却液隔开,两边的液体不相混。Roof chillers supply the primary chilled water. Inside the CDU it stays apart from the secondary coolant. The two liquids never mix.效果图_06_屋面冷源.jpg海湾夏天按室外五十度留裕量。佳力图已交付项目的年化 PUE 做到一点三到一点四,那是交付实绩,不是这座园区的承诺。A Gulf summer is designed with margin for fifty degrees outside. On projects Canatal has already delivered, annualised PUE reached 1.3 to 1.4. That is a record of delivery, not a promise for this campus.效果图_21_屋面热浪.jpg
5. 预制电力模块5. Prefabricated power module
要盯住的风险不在数据大厅,而在预制电力模块的电池室门口。The risk to watch is not in the data hall. It is at the battery-room door of the prefabricated power module.效果图_22_模块门口.jpg模块在工厂里集成、测试,再运到现场就位。The module is integrated and tested in the factory, then set in place on site.效果图_12_工厂预制.jpg门外是变压器、精密空调室外机、电池室的门,以及预留的消防水接口。一个模块大约八间电池室,每间约四百八十五千瓦时,合计约三点九兆瓦时。Outside: the transformer, the precision-cooling condensers, the battery-room door, and a fire-water inlet left ready. About eight battery rooms, about 485 kilowatt-hours each, about 3.9 megawatt-hours together.效果图_08_预制电力模块.jpg佳力图配大约十台精密空调,把电池留在合适的温度和湿度里。温度异常时,先由精密空调加力。Canatal fits about ten precision air conditioners, and holds the batteries at the right temperature and humidity. If the temperature rises out of line, they cool harder first.效果图_24_精密空调.jpg
6. 主动安全与极柱液冷6. Active safety and terminal cooling
舱内是祎增的高倍率磷酸铁锂,支撑 UPS 大约十分钟。顶棚水淋只是待命,画面里没有喷水。Inside is Yizeng high-rate lithium iron phosphate, enough for about ten minutes of UPS backup. The spray overhead is only standing by. Nothing in the picture is spraying.效果图_03_电力模块安全储能舱.jpg喷头处于待命。电芯到了六十三度才淋水,赶在失控之前。The nozzles are standing by. Water cools the cells only at 63 degrees, before runaway.效果图_11_水淋喷头.jpg角落是自带水箱,墙上是市政消防水接口。气体可以扑灭明火,却带不走电芯里面的热。A tank in the corner, a municipal fire-water connection on the wall. Gas can put out a flame, but it cannot take the heat out of the cell.效果图_23_水箱接口.jpg每天真正控温的是极柱液冷。液冷板贴在极柱上,电芯之间的温差就小了。What holds the temperature every day is cooling at the cell terminals. The plate sits on the terminals, so the difference between cells stays small.效果图_07_极柱液冷.jpg极柱热路示意:热从极柱进入液冷板。水淋要到六十三度以后才用。The heat path at the terminal: heat enters the cold plate from the post. The spray is used only after 63 degrees.效果图_31_极柱热路.jpg
7. 园区储能7. Campus energy storage
园区这一层按项目来选。直流柜平滑训练负载的秒级波动,也填上柴油发电机启动前的空档,一个模块大约十兆瓦时。The campus layer is chosen project by project. DC cabinets smooth the second-by-second swing of a training load, and fill the gap before the diesel generators start. About 10 megawatt-hours for one module.效果图_04_园区储能.jpg屋面和停车场的光伏配上储能,绿电就能更多地留在现场。Roof and car-park solar, with storage, keep more of the green power on site.效果图_25_光伏车棚.jpg施工和联调用一点九到四点九兆瓦时的移动储能,替代临时柴油,也可以作为负载测试的电源。During construction and commissioning, mobile storage of 1.9 to 4.9 megawatt-hours replaces temporary diesel, and it can also power a load test.效果图_26_移动储能.jpg
8. 一个界面,仿真与监控8. One interface, simulation and monitoring
佳力图与祎增把冷却和用电安全收在同一个界面上。Canatal and Yizeng bring cooling, and the safety of the power, onto one interface.效果图_17_一个界面.jpg孪生对照是示意,不是已经交付的软件界面。The twin is a comparison view. It is not a software screen already delivered.效果图_29_孪生对照.jpg算力在液冷机房,储能在电池间。仿真屏把散热、电力和水淋策略放在一起对照。Compute in the liquid-cooled hall, storage in the battery room. The simulation screen sets cooling, power and the spray strategy side by side.效果图_27_仿真大屏.jpg冷板、管路和储能柜按供液和回液温度着色,用来看热走到了哪里。Cold plates, piping and storage cabinets are coloured by supply and return temperature, so the path of the heat can be seen.效果图_28_数字孪生.jpg电芯温度、供水流量和储能功率一路被看着,汇进同一块数字大屏。业主看见的仍然是一个界面。Cell temperature, water flow and storage power are watched all the way through, and they meet on one digital wall. What the owner sees is still one interface.效果图_30_监控大屏.jpg