AI-POD-1 · 4 × 80 kW GPU racks · direct-to-chip · 400 kW CDU · roof dry cooler
AI pod digital twin
| Point | Now | Expected | Status | Low | High |
|---|
How the twin judges health
Every reading is compared with what a healthy pod would show at the same rack power and outdoor air, using the physics the drawing runs on: the cold plates' heat from flow and rise, the plate exchanger's approach at the measured flows, the pump's curve and the dry cooler's fans. Each check uses the measured neighbours, so a fault lights only its own part. A reading that stops arriving is No data, never normal, and data older than 3 minutes is a lost-communication alarm. pPUE here covers the liquid cooling chain: CDU pumps, facility pump and dry-cooler fans.
Connect real data
Open the page with ?feed=https://your-gateway/latest.json and the twin polls that URL every 5 s instead of running the simulator. It expects one reading, or a list, with these fields:
t(ISO time or epoch ms),ambient_c,it_kw,cooling_kw- per rack
rackN_kw,rackN_flow_ls,rackN_return_c,rackN_gpu_c(N = 1–4) tcs_supply_c,tcs_return_c,tcs_flow_ls,cdu_dp_kpa,tcs_pressure_kpa,exp_tank_pct,strainer_dp_kpapump1_speed_pct,pump1_kw,pump1_vib_mm_s, and the same forpump2fw_valve_pct,fw_supply_c,fw_return_c,fw_flow_lsdc_fan_cmd_pct,dc_fan1_rpm…dc_fan4_rpm,dc_fan_kw
About this pod
A generic pod: four racks of GPU nodes at 80 kW each, cooled direct-to-chip with PG25 at 35 °C supply, 85% of the heat to liquid. One 400 kW CDU pumps the secondary loop (duty and standby pumps, plate heat exchanger, strainer, expansion tank, flow meter) and hands the heat to a glycol facility loop and a dry cooler with four EC fans on the roof deck, drawn closer than its true height. Numbers come from a simple steady-state model and a simulator; they are illustrative, not from a site.