PDF 原檔:報告_OCP2026_NVIDIA_VeraRubin機櫃功率平滑_20261015_original.pdf
講者公開簡報(OCP 官方 Recorded Content 頁提供下載)。文字以 PyMuPDF 逐頁抽取,
### p<n>為頁碼;圖表內容見原檔。
圖片清單(已驗證 2026-10-19)
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_NVIDIA_VeraRubin機櫃功率平滑_20261015_p2.png |
p2 | 真資料圖 | 功率曲線示意:無平滑(黑線,啟動尖峰與週期性大幅掉落)vs 有平滑(綠線,受控 ramp-up、平穩 steady-state、ramp-down),標示 Power Cap、GPU Burn、Energy Storage |
報告_OCP2026_NVIDIA_VeraRubin機櫃功率平滑_20261015_p3.png |
p3 | 真資料圖 | Vera Rubin 機櫃實物圖,旁附 GPU 板與機櫃內儲能托盤(托盤內可見成排圓柱形元件);左側文字為 400 J/GPU 機櫃內儲能、多組閉迴路控制器跨托盤通訊、充電狀態依 AC 爬升率等條件 |
報告_OCP2026_NVIDIA_VeraRubin機櫃功率平滑_20261015_p4.png |
p4 | 真資料圖 | LLM 訓練實測 #1:AC whip 功率 20 秒時序,IPS OFF 約 40–45kVA 並週期性掉到約 14kVA,IPS ON 平貼約 40kVA;標註 13% 峰值降低 |
報告_OCP2026_NVIDIA_VeraRubin機櫃功率平滑_20261015_p5.png |
p5 | 真資料圖 | LLM 訓練實測 #2:IPS OFF 約每秒一次由約 15–20kVA 低點衝到近 60kVA 尖峰,且 15–17.5 秒有約 14kVA 低谷,IPS ON 平貼約 38–39kVA;標註 51% 峰值降低 |
原始內容
p1
Gabriele Gorla, Rouslan Dimitrov NVIDIA Intelligent Power Smoothing at the Vera-Rubin Rack Market Impact & Adoption ]
p2
Intelligent Power Smoothing Reclaims Squandered Power Ramp-down Power Time Ramp-up Steady-state Power profile without smoothing Power profile with smoothing Energy Storage Power Cap Idle GPU Burn Make the rack an ideal grid load: controlled ramp up and ramp down smooth steady state Decrease the power peaks for more tokens per provisioned power
p3
Energy Storage and GPUs Work Together 400 J / GPU of energy storage in the rack Multiple closed-loop controllers communicate across trays State of charge level dependent on: allowed AC ramp rate workload power profile GPU efficiency optimizing controller
p4
LLM Training Result #1 LLM training, Vera-Rubin NVL72 rack 13% peak power reduction AC whip power
p5
LLM Training Result #2 LLM training, Vera-Rubin NVL72 rack 51% peak power reduction AC whip power
p6
IPS Hardware Requirements NVIDIA Approved Vendor List (AVL) Power Shelves: 110 kW AC/DC Power Shelves 800V Power Rack + 90kW DC/DC Power Shelves Compute Trays: Rubin GPU equipped systems (NVL72 and future) More vendors under final validation
p7
Single script configures the entire rack (power shelves, HPMs, and GPUs). Script and documentation at nvonline ID 1153760. IPS Software Enablement Configure — config.json Commands Add power-shelf PMC and compute BMC addresses. Choose the IPS settings below. "power_smoothing": { "psu_max_ac_ramp_rate_w_per_s": 400, "gpu_ramp_up_w_per_s": 20, "gpu_ramp_down_w_per_s": 20, "gpu_idle_hysteresis_seconds": 10 } Apply settings and verify readback: ./ips_power_smoothing.sh apply config.json Check current settings: ./ips_power_smoothing.sh status config.json Disable IPS: ./ips_power_smoothing.sh disable config.json
p8
Open Discussion