PDF 原檔：報告_OCP2026_NVIDIA_DSX_MaxLPS_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_DSX_MaxLPS_20261015_p4.png |
p4 | 真資料圖 | MaxP vs MaxLPS 對照表：GPU TGP 1,200W vs 800W、機櫃供電 120kW vs 90kW、吞吐相同、同電力包絡可多 33% 機櫃（GB200，Kimi-K2.5 FP4） |
報告_OCP2026_NVIDIA_DSX_MaxLPS_20261015_p5.png |
p5 | 真資料圖 | 靜態分配 vs 動態分配長條圖：540kW 預算下 4 櫃（GB300 實測 95／65／90／80kW）消耗 330kW、閒置 210kW；MaxLPS 動態分配消耗 425kW 並多放 1 櫃（DeepSeek-R1 95kW） |
報告_OCP2026_NVIDIA_DSX_MaxLPS_20261015_p6.png |
p6 | 真資料圖 | MaxLPS 架構圖：Dynamic Power Software 與 Workload Optimizer 經 OOB 設定 GPU／CPU／NIC／交換器功率上限，向上接 DSX Exchange、DSX Flex（MQTT）、電網 ISV／EmeraldAI、Phaidra AI Agents、BMS，下接 PQM／RPP／UPS／CDU 遙測 |
報告_OCP2026_NVIDIA_DSX_MaxLPS_20261015_p7.png |
p7 | 真資料圖 | 閉環控制與逐節點決策流程：消耗 vs 配置功率，差距 >20% 降配、5–20% 不動作、近滿載（5% 內）加配，更新可用 DC 功率後連續迴圈 |
報告_OCP2026_NVIDIA_DSX_MaxLPS_20261015_p8.png |
p8 | 真資料圖 | 優先級合規長條圖：P0（Max-P）／P1／P2（Max-Q）工作負載功率與 50kW floor，事件時依序 P2、P1、P0 降至 floor，最後才停工作，事件結束恢復；另示意多出一櫃 110kW |
原始內容
p1
Ajay Kamalvanshi Senior Director, NVIDIA Every Megawatt Counts: The Path to Building AI Factories with DSX MaxLPS
p2
• Background and Problem Statement • Architecture and Theory of Operation • DSX Stack Integration • Ecosystem Partners • Challenges and Collaboration Agenda
p3
AI Factory Output Is Constrained by Power Energy is at the foundation of the 5-layer AI cake: power constraints limit growth Measured GPU Power Across AI Workloads on GB300 Inference workloads, no power capping Workload power needs vary; not all workloads need Max-P power Static provisioning strands power There is so much power that is squandered away in AI Factories… every unused watt is revenue lost – Jensen GTC 26 keynote Example
p4
Problem: Power is not optimally utilized Shows 33% more GB200 racks based on Kimi-K2.5 FP4 Dynamo TRT-LLM 8K/1K Rack Power MaxP MaxLPS GPU TGP setting 1,200W 800W Provisioned Rack Power 120kW 90kW Throughput 1x Same Racks in Same Power Envelope 1x 33% More Racks* * 120kW / 90kW = 1.33x
p5
5 Static power allocation: Total budget = 540 kW, Consumed = 330 kW | Stranded = 210 kW 95 kW Deepseek-R1 65 kW 70 kW 80 kW 55 kW 95 kW 40 kW No power budget for more rack Workload power consumption Power allocated but unused Legend 90 kW 45 kW DeepSeek-R1 Qwen3-235B GPT3-175B Mixtral-8x7B 65 kW 80 kW 95 kW 90 kW DeepSeek-R1 Qwen3-235B GPT3-175B Mixtral-8x7B Measured GB300 power Additional rack MaxLPS dynamic allocation: Same 540 kW budget Consumed = 425 kW | +1 additional rack enabled Allocated Rack Power MaxLPS recovers stranded power Deploying additional GPUs within the same power budget
p6
MaxLPS Architecture Dynamic Power Software (DPS), Workload Profiles, and Optimization Policies Application Dynamic Power Software DSX Exchange Grid ISV/EmeraldAI Topology/Events Power Shelf Node GPU CPU Switches OOB to set limits Power telemetry DSX Flex (MQTT) Zapp PQM, RPP, UPS, CDU NIC AI Agents (Phaidra AI) BMS MQTT MQTT Close Loop Control Workload Optimizer • Maximize tokens/watt with dynamic power and performance optimizations. • Enable dynamic power management based on workloads and optimization goals Max-P: maximizing performance to meet SLA MaxLPS: maximizing node count with improved perf/provisioned-watt • Unlock stranded utility power (100GW in North America) with ISV like Emerald AI • Workload Optimizers: • Workload Profile (memory clock, switch optimization, process optimization) • Utilization/observed power-based prediction • Heuristic/DL based prediction MaxLPS
p7
7 Closed-Loop Control - Details Closed-Loop Control Architecture Power Provisioning Power Telemetry Closed Loop DPS MaxLPS Decision Algorithm (per Node) START, for each node Consumed power vs allocated power END Reduce allocation to node Increase allocation to node No action Update available DC power much less (>20% gap) near capacity (within 5%) on target (5-20% gap) Continuous Loop Guarantees compliance within total DC power budget | Simple algorithm ensures robustness
p8
110KW Additional rack Power Compliance with Workload Priority Handling SLA and power events like Load Shedding Workload priorities are set along with cap and floor power as part of power policy setting Floor* = 50 kW Event P2 WLs throttle to floor If compliance not met, P1 WLs throttle to floor If compliance not met, P0 WLs throttle to floor (If needed) Stop jobs (last resort) Power restored when event ends 105KW 109KW 79KW 87KW 95KW P0 priority workloads – Max-P provisioning P1 priority workloads provisioned to workload need, prioritized for boost P2 priority workloads – Max-Q provisioning, lower priority for boost 25KW 43KW
p9
DSX Stack Integration for Vera Rubin Blueprint for scaling MaxLPS to 1GW Datacenter • Scaling for large-scale clusters using DSX stack • Integration with Provisioning systems: NICo, and Launch Layer • Supporting multi-tenants through DSX’s BGP, EVPN, and VXLAN technologies • OT interfaces through DSX Exchange • Include NVLink Switch and LPU MaxLPS to improve throughput for premium and ultra segments
p10
Lambda reclaiming stranded power to increase compute density within the same infrastructure footprint. AI Clouds Deploy NVIDIA DSX, Lambda Adopts MaxLPS “NVIDIA DSX MaxLPS paves the way to reclaiming stranded capacity and converting it into real-world usage, with significantly more compute density in the same footprint. It’s a scalable and sustainable blueprint for how we can operate and expand our infrastructure moving forward.” —Dave Ward, President, Cloud Services, Lambda 네이버클라우드기업정보 - 2026 년연봉 , 사업 , 채용 , 지원자스펙분석
p11
Challenges & Collaboration • Improve latency and reliability of BMC Redfish interfaces with IBV/IFVs • Support fast streaming telemetry to scale across the datacenter – move from Modbus/Redfish to UDP-based streaming interfaces in collaboration with the OCP GPU Management Interfaces workgroup • NVIDIA Zapp as a reference interface implementation from NV for v2.0.0 spec • Incorporate fast streaming telemetry on power devices like PQMs and BESS for communication with the system management stack • DSX Stack as a reference stack for AI factory – simulate with DSX Sim Moving towards no stranded power