PDF 原檔：報告_OCP2026_Denvr_UNICOM_浸沒原生IT設備部署_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 2026-10-15）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_Denvr_UNICOM_浸沒原生IT設備部署_20261015_p4.png |
p4 | 真資料圖 | Denvr 模組化資料中心對比：氣冷 Co-Lo 1 MW／6400 sq ft vs Denvr 模組 1 MW／600 sq ft 空拍示意；128 GPU／900 sq ft、零水 PUE<1.04、6–12 個月部署；3OU H200 SXM 伺服器每櫃 14 台安裝照 |
報告_OCP2026_Denvr_UNICOM_浸沒原生IT設備部署_20261015_p6.png |
p6 | 真資料圖 | 浸沒最佳化 SXM 伺服器 vs 氣冷 XE9680 實體照：體積縮 2 倍、每櫃伺服器 4 倍（14 顆 H200）、每台省約 800 W |
報告_OCP2026_Denvr_UNICOM_浸沒原生IT設備部署_20261015_p12.png |
p12 | 真資料圖 | H200 SXM 叢集壓力測試時序圖：槽體負載（約 80 kW 階段）、流量（約 70 gpm）、進出水溫（進約 35°C、出 43–47°C）及五日三階段變化 |
報告_OCP2026_Denvr_UNICOM_浸沒原生IT設備部署_20261015_p14.png |
p14 | 真資料圖 | 1000W+ GPU 浸沒改造：銅製浸沒最佳化散熱器伺服器內部照與 Direct Flow Manifold（四組泵閥）3D 圖，N+1 冗餘、伺服器管理流量、冷卻功耗 12.5–75 W |
報告_OCP2026_Denvr_UNICOM_浸沒原生IT設備部署_20261015_p16.png |
p16 | 真資料圖 | 同平台 8×SXM GPU 1,000 W 目標 600 秒 DCGM 測試，浸沒（35°C）vs DLC（30°C PG25）：平均功率 996／995 W、峰值 Tcore 77／64°C、8/8 通過 |
原始內容
p1
Amy Short, PhD Executive Officer – Advanced Data Center Technologies, Denvr Dataworks Austin Hipes Chief Technologist, VP of Engineering, UNICOM Engineering Deploying Immersion Native ITE in AI Neocloud: An OCP Collaboration Story Cooling Environments Immersion Project
p2
Speaker Introductions & Tech Deep Dive QR Amy Short, PhD Exec. Officer – Advanced Data Center Technologies Denvr Dataworks Austin Hipes Chief Technologist & VP, Engineering UNICOM Engineering "After Hours" Technical Deep Dive Download our Global Summit slides, whitepapers, and podcast episodes
p3
Technical & Commercial Challenges • Power & thermal constraints • Water & sustainability needs • Deployment speed • Warranty and long-term reliability Why We’re Here Today • Few production examples for high density GPU platforms • To prove immersion is scalable & performant – we have not yet seen a ceiling • Solves a real COMMERCIAL challenge – it is not a lab experiment • Clear demand for differentiated AI infrastructure Challenges in AI Deployments
p4
• Modular Standardized and repeatable • High Density 128 GPUs in 900 sq ft • Sustainable Zero-water, PUE <1.04 • Built for rapid scale 6-12 months to deploy Denvr Approach to Infrastructure at Scale Air Cooled Co-Lo 1 MW per 6400 sq ft Denvr Dataworks Modular Data Center 1 MW per 600 sq ft 3OU H200 SXM servers deployed 14 per rack at Denvr
p5
UNICOM Engineering Component-level and server-level validation Denvr Dataworks Server-level and cluster-level validation; commissioning and deployment Collaborative AI Server Development Program Complex challenges demand collaboration and complementary capabilities
p6
An OCP-Aligned foundation establishes a standardized approach to address power, thermal, density, and deployment challenges Immersion-Optimized SXM Server Design Air Cooled XE9680 Server Design H200 SXM server Denvr Immersion Optimized Server Design H200 SXM server 2x server size reduction 4x servers per rack (14x H200s) ~800 W per server power savings
p7
Warranty Coverage Based on 2.5 Years of Reliability Testing New Physical Design • Reduced physical footprint by 2x (6U → 3OU form factor) • Based on the Dell XE9680 Platform for H100 and H200 GPUs Power Architecture Updates • 102 kW power savings per 128 server scalable unit • Inspired by OCP Open Rack designs • Adaptable for ORV2.2 and ORV3 Thermal Redesign • Immersion-ready copper heat sinks • Evaluated in system temperatures up to 60 °C Software & Firmware Optimization • Updated firmware • Co-development with Dell Rethinking Server Design for AI Use Cases Fourteen H200 SXM servers with 3OU chassis supported in a single Denvr production environment tank.
p8
Thermals, not power, limit density in air-cooled racks
• Immersion increases usable compute per square foot
Stable thermals improve reliability and warranty confidence
• Reduced CPU/GPU variability
• Validated via NVQual and production benchmarking
Thermal Performance
Full data set available on Denvr-UNICOM Engineering “After Hours” landing page linked at end of presentation
NVQual Thermal Test Result
Test Application
TDP
Test Result
Pass
Test Result (Immersion)
Parameter
GPU<5d.00.0>
GPU<3b.00.0>
GPU
p9
Signal Integrity must be designed and validated for Immersion • Not assumed from air-cooled architectures Data-driven Validation • NVQual testing confirms signal integrity • Benchmarking shows stable, sustained performance Signal Integrity Signal Integrity maintained in immersion Full data set available on Denvr-UNICOM Engineering “After Hours” landing page linked at end of presentation
p10
Hardware: 8x GPU H200-SXM5-141GB Denvr MSC1 GPU TDP: 700 W Test Type: Continuously batched LLM inferencing of Llama-3.1-Instruct 70B Details: • Tensor Parallelism across all 8X H200 accelerators • Context length per batched sequence: 4096 toks (3072tok/prompt, 1024tok/response) • Maximum of 150 concurrent requests • OTPS (output toks per second): 8068 • RPS (responses per second): 7.9 • P99 per-response latency (1024 toks): 19.82s Demonstrated In Real AI Workloads 500W 520W 540W 560W 580W 600W 620W 640W 660W 680W GPU Power Time Test data by Denvr; Full data set available on Denvr-UNICOM Engineering “After Hours” landing page linked at end of presentation
p11
Denvr Infra Response to Multi-Day AI Stress Test Phase 1 100% of servers in high stress HPL testing Phase 2 75% of servers HPL, 25% bursting OEM testing Phase 3 25% server HPL 25% intermittent OEM testing 50% idle Representative example from Denvr H200 SXM cluster commissioning Immersion Tank Load Real Power (kW)
p12
Denvr Infra Response to Immersion H200 SXM Stress Test: Reliability, Responsiveness & Scalability Power Draw • Measured IT Load responds as expected. • ~800 W per server power savings vs. air cooled parent configuration. Flow Rate • Increases as expected in response to IT load. • Stabilizes quickly in each test phase. Mild Temperatures, Longer Lifetime Servers • Peak system temp increases only 6°C at and quickly stabilizes at ~45 °C. • Rack inlet/outlet ΔT ~10°C maintained as targeted. • FMEA testing demonstrated 15 min-1 hr response time possible due to mild env. Representative example from Denvr H200 SXM cluster commissioning Immersion Tank Load Real Power (kW) Immersion Tank Flow Rate (gpm) Immersion Tank Inlet Temp (°C) Immersion Tank Outlet Temp (°C) Min: 43 ℃ Max: 47 ℃
p13
1,000 W+ Target power per GPU Targeted Flow 3000W+ Per OU IMMERSION-NATIVE AI INFRASTRUCTURE Design for the next thermal envelope Immersion-optimized heatsinks, direct flow delivery, redundancy, and server-managed controls create a path beyond today’s Blackwell validation point. Immersion server architecture validated for sustained 1,000W+ class GPU operation, PCIe/NVLink performance, and scalable thermal management requires additional enhancements. Design Updates For Blackwell Generation GPUs & Beyond
p14
Server-level thermal hardware, redundant flow delivery, and closed-loop management work together as one cooling architecture Immersion-Optimized Heatsinks Installed Patent-Pending DFM Installed Inside the Server Cold plates replaced Immersion-optimized copper heatsinks increase wetted surface area for direct heat transfer. Direct Flow Manifold Integrated, low-profile DFM hardware delivers coolant to critical components. N+1 redundancy Redundant flow paths maintain enhanced cooling capability. Server-managed control System management adjusts flow to individual server thermal needs. Energy-aware operation Flow and pumping power are optimized together. Cooling power consumption ranges between 12.5W – 75W total. Immersion optimized design: optimized heatsinks transfer heat to the fluid, while patent-pending DFM hardware and server system management deliver redundant, demand-based flow with lower unnecessary pumping power. Physical Modifications Enhance Immersion Cooling for 1000W+ GPUs
p15
8 SXM GPUs • 1,000 W target per GPU • 600-second DCGM targeted-power test • 35°C Dielectric Fluid Inlet 996 W Mean GPU power 0.4% Below 1,000 W target 77°C Peak GPU core 83°C Core thermal limit 8 / 8 DCGM pass Power and clock remain tightly grouped while core temperatures stay below the stated limit. Power Mean sustained power during active test Clock Average GPU clock during active test GPU Core Temperature Maximum measured Tcore 1,000 W GPU 0 994 W GPU 1 997 W GPU 2 996 W GPU 3 997 W GPU 4 996 W GPU 5 996 W GPU 6 996 W GPU 7 996 W Across GPUs 0–7 0 1 2 3 4 5 6 7 990W 1000 W GPU 0 1595 MHz GPU 1 1536 MHz GPU 2 1636 MHz GPU 3 1532 MHz GPU 4 1549 MHz GPU 5 1557 MHz GPU 6 1595 MHz GPU 7 1539 MHz Across GPUs 0–7 0 1 2 3 4 5 6 7 1500 MHz 1660 MHz 83°C limit GPU 0 69°C GPU 1 77°C GPU 2 68°C GPU 3 76°C GPU 4 77°C GPU 5 68°C GPU 6 68°C GPU 7 77°C Across GPUs 0–7 0 1 2 3 4 5 6 7 65°C 85°C Takeaway: The 8-GPU system sustained 99.6% of target power while peak core temperature remained 6°C below the stated limit. Blackwell SXM Sustains Full Power Under Load 15
p16
Immersion, 35°C DLC, 30°C PG25 Mean power 996 W 995 W Mean clock 1,567 MHz 1,594 MHz Peak Tcore 77°C 64°C DCGM result 8 / 8 PASS 8 / 8 PASS Power Mean sustained power by GPU 1000 W 0 1 2 3 4 5 6 7 990 W 1000 W GPU index Immersion 994.5–997.3 W DLC 992.8–997.6 W Clock Average GPU clock during load 0 1 2 3 4 5 6 7 1500 MHz 1650 MHz GPU index Immersion 1532–1636 MHz DLC 1568–1618 MHz GPU Core Temperature Maximum measured Tcore 83°C 0 1 2 3 4 5 6 7 50°C 85°C GPU index Immersion 68–77°C DLC 55–64°C Takeaway: Both cooling approaches passed all 8 GPUs and sustained at least 99.5% of the 1,000 W target on average. Immersion matched DLC power delivery while remaining below the 83°C Tcore limit. Same platform • 8 × SXM GPUs • 1,000 W target per GPU • 600-second DCGM targeted-power tests Delivers Comparable Full Power Performance in Immersion & DLC 16
p17
Immersion 090534 DLC 000332 PCIe CPU → GPU Immersion 55.34 GB/s DLC 55.34 GB/s +0.00% vs DLC PCIe GPU → CPU Immersion 57.87 GB/s DLC 57.14 GB/s +1.28% vs DLC NVLink GPU → GPU Immersion 766.32 GB/s DLC 767.46 GB/s -0.15% vs DLC NVLink GPU ← GPU Immersion 774.49 GB/s DLC 774.56 GB/s -0.01% vs DLC NVLink bidirectional read Immersion 1525.27 GB/s peak DLC 1526.62 GB/s peak -0.09% vs DLC NVLink bidirectional write Immersion 1540.31 GB/s peak DLC 1540.24 GB/s peak +0.00% vs DLC Result: All six bandwidth tests passed in both cooling configurations. Immersion remained within 0.15% of DLC on directional NVLink and within 1.27% on directional PCIe. PCIe replay events 0 EDATL error events 0 Immersion samples 28,360 Like-for-like DCGMI NVBandwidth results • CPU GPU represents PCIe • GPU GPU represents NVLink Immersion Matches DLC PCIe & NVLink Bandwidth Takeaway: Immersion bandwidth matched the DLC baseline, while separate monitoring observed zero PCIe replay-counter events and zero EDATL error events.
p18
• This work was enabled by OCP: • Power distribution & open rack designs • Immersion requirements • Material compatibility • Signal Integrity • Safe handling and cleaning • FMEA and operational guidance • Engage with OCP technical committees to help shape what comes next! Call to Action: Enable Innovation Through Collab! Register and submit your questions in advance to help shape the conversation. Slides and Technical Deep Dive For deeper insights on the topic, scan the QR code here! After Hours discussion – Ask Me Anything!!
p19
Thank You! Access these slides and other technical deep dive content. Schedule a meeting with our team. After-hours discussion topics – Ask Me Anything!