PDF 原檔：報告_OCP2026_TE_448G高速IO直液冷_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_TE_448G高速IO直液冷_20261015_p3.png |
p3 | 真資料圖 | 可插拔 IO 模組 TDP 對 SerDes 速率（56G→448G，對數軸）：氣冷上限約 35W，448G 預測 8 線約 70W、16 線約 120W、32 線約 250W、64 線約 500W；224G 的 XPO 64 線標在約 250W 帶（模組內液冷） |
報告_OCP2026_TE_448G高速IO直液冷_20261015_p4.png |
p4 | 真資料圖 | 為何 IO 要液冷：每線功率 56G→448G 增 5 倍（1.5→7.5W）；長條圖 56G 8 線 12W、112G 16W、224G 27W（已液冷出貨）、448G 8/16/64 線預測 70/120/500W；紅虛線 35W 氣冷上限；DSP 熱通量 100+ W/cm2 |
報告_OCP2026_TE_448G高速IO直液冷_20261015_p7.png |
p7 | 真資料圖 | 448G 直接晶片冷板兩種設計：8/16 線整合冷板模組（85x121mm，2 個熱源 70/120W）、32/64 線 5mm 微通道冷板（72x30mm，4 或 8 個熱源，250/500W），冷卻液 PG25 |
報告_OCP2026_TE_448G高速IO直液冷_20261015_p10.png |
p10 | 真資料圖 | 模擬結果：流量對最高溫曲線，64/32/16/8 線全部低於 95C 上限，右側標示最高晶片溫 <85C、進水 45C |
報告_OCP2026_TE_448G高速IO直液冷_20261015_p13.png |
p13 | 真資料圖 | 實測 vs 模擬箱形圖：250W/500W、0.7/0.9 lpm 下最高溫約 61-75C（95C 上限）；壓降約 0.8-1.25 psi，實測與模擬貼近 |
原始內容
p1
Divya Bandaru, System Architect, TE Connectivity Sridhar Mukthineni, Sr Mgr. ME COE, TE Connectivity Arjun P, R&D DVL engineer, TE connectivity Remesh Kuzhikkali, Principal R&D DVL engineer, TE Connectivity Cooling the 448G Frontier: Thermal Architecture for 500 W-Class HSIO Using Direct-to-Chip Liquid Cooling COOLING: COLD PLATE
p2
Agenda 01 Introduction Pluggable I/O module: Thermal Design Power trends 02 System Architecture Direct-to-chip vs. External cold plates 03 Direct to Chip Cold plate designs for 448G IOs 04 Future Work Next steps and call to action
p3
Introduction: Pluggable I/O Thermal Evolution SerDes lane rate (Gb/s per lane) IN-MARKET TRENDS • 4-lane: QSFP56 → QSFP112 • 8-lane: OSFP 400G → 1.6T RHS • 64-lane: XPO for 224G (in- module liquid) 16-lane ~120 W 64-lane ~500 W 5 10 30 100 300 1000 I/O module TDP (W) [Log scale] 56G 112G 224G 448G (TDPs are projected) Air-cooled riding heat sink: ~25–35 W In-module direct-to-chip liquid (quick-disconnect): ~120- 500 W Air cooling limit QSFP56 QSFP112 QSFP-DD / OSFP 400G 800G OSFP / QSFP-DD OSFP-RHS 1.6T XPO 64-lane (in-module liquid) 4-lane 8-lane 16-lane 64-lane 8-lane ~70 W 32-lane ~250 W Liquid Cooling beyond > 35w • Cage cold plate with thermal bridge (dry TIM)→ ~50 W. • Liquid cooling→ 35-500 W. External cold plates/Riding heat sinks attached to cage ~35–50 W 16 lane 224G 3.2T
p4
Why do I/O Modules need liquid cooling? Power per lane Increased 5x for 56G→448G (1.5 → 7.5 W) Air- Cooling limitations Highest air-cooled module In market is ~35W Heat Flux 100+ W/cm² for DSP chips (> most CPUs) 19″ Standard racks Same rack, same 32 ports for 1RU 0.4 kW @ 56G → 0.9 kW @224G → 1.9 kW per RU for faceplate 12 W 56G (8 Lane) 16 W 112G (8 Lane) 27 W 224G (8 Lane) 70 W 448G 8 Lane (projected) 120 W 448G 16 Lane (projected) 500 W 448G 64 Lane (projected)
35 W — no module above this line has ever shipped on air already ships liquid-cooled Physics of air cooling met its limitation for beyond 1.6T, 224G Modules
p5
Why Direct-to-Chip is the Path for Future I/O EXTERNAL — COLD PLATE IN CONNECTOR CAGE External cold plate- RECP Dry TIM / Thermal Bridge — RTB Module top — (RTop) TIM1 (RTH ) TIM1 (RTH ) DSP TIA / driver Paddle PCB ΣR = RTH + RTop + RTB + RECP MOVE THE WATER INSIDE MODULE: ⟶ DIRECT-TO-CHIP : IN-MODULE INTEGRATED COLD PLATE — RICP (blind-mate QDs) TIM (Rth) TIM (Rth) DSP TIA / driver Paddle PCB ΣR = RTh + RICP Four layers in the stack vs. Two → Fewer components, lower ΣR. Every layer between the die and coolant adds to thermal resistance stack up Module top
p6
External Cold Plate vs In-Module Direct-to-Chip at 448G EXTERNAL — CAGE COLD PLATE ✓ No leakage risks ✓ Widely adopted, lower risk of infant failure rate ✓ Aids dry hot-swap for pluggable IO ✗ Current dry TIMs have high thermal resistances ✗ Some dry TIMs age every insertion cycle (max mating 10 cycles). ✗ Dry TIM requires very high compressive forces ✗ Proven compression hardware to support TIM compression ✗ Challenging to cool high-density modules IN-MODULE — DIRECT-TO-CHIP ✓ Ability to cool with margins at 500W ✓ Better thermal performance = higher reliability for module components ✓ Can meet 100 insertion cycles ✗ Ecosystem is immature, higher chance of infant mortality rate. ✗ Leakage risks, interop challenges between QD vendors. ✗ High pressure drops with mini QDs. vs
p7
448G: DTC Cold Plate Designs Two cold plate concepts evaluated for thermal and pressure drop performance across 8, 16, 32 and 64 lanes 8 & 16 LANES — INTEGRATED COLD PLATE MODULE · 70 / 120 W PG25 fluid inlet/outlet lines Cold plate / cooling channels integrated into module 2 heat sources – 70 W for 8 lane, 120 W for 16 lane 32 & 64 LANES — 5 MM MICROCHANNEL COLD PLATE · 250 / 500 W 32-lane concept – 4 heat sources (250 W), 4 on top Cold plate lid Cold plate body 64-lane concept – 8 heat sources (500 W), 4 on top, 4 on bottom PG25 fluid inlet/outlet lines Representative heat sources
p8
POWER PER CONCEPT — DISTRIBUTED EQUALLY ACROSS HEAT SOURCES x8 70 W 2 heat sources x16 120 W 2 heat sources x32 250 W 4 heat sources x64 500 W 8 sources – 4 top, 4 bottom MODEL ASSUMPTIONS — THERMAL, FLOW & BOUNDARIES TIM RESISTANCE 0.05 °C/W – worst case for PTM-type material MATERIAL & FLUID Copper cold plate and lid, PG 25 coolant INLET TEMPERATURE System and fluid inlet at 45 °C HEAT TRANSFER All chip heat removed entirely by the coolant FLOW RATES 1.5 lpm per KW – 1. 0.1/ 0.2 / 0.3 lpm for x8 & x16 2. 0.4/ 0.7 / 0.9 lpm for x32 & x64 BOUNDARY All walls adiabatic – no heat loss to surroundings 448G: DTC Cold Plate Design - Assumptions Two cold plate concepts evaluated for thermal and pressure drop performance across 8, 16, 32 and 64 lanes
p9
MAX CHIP TEMP 74.9 °C Target < 95 °C • fluid 45 → 55 °C 448G: DTC Cold Plate Simulation Results TOP VIEW BOTTOM VIEW 64-lane concept at 500 W and 0.7 lpm — steady-state temperature distribution Fluid inlet Fluid outlet Section X-X Maximum temp = 74.9 °C Maximum temp = 74.9 °C Section Y-Y Section Y-Y Section through Sources 2,3,6,7 Section X-X Section through Sources 1,4,5,8
p10
MAX CHIP TEMP < 85 °C across every scenario evaluated ALLOWABLE LIMIT 95 °C dashed line on chart INLET TEMPERATURE 45 °C system and fluid at start 448G: DTC Cold Plate Simulation Results All cold plate scenarios evaluated for thermal and pressure drop performance THERMAL PERFORMANCE — ALL SCENARIOS MEET THE TARGET
p11
x32 / x64 COLD PLATE < 2 psi at 0.7 – 0.9 lpm x8 / x16 COLD PLATE < 9 psi at 0.1 – 0.2 lpm MINI QUICK DISCONNECTS 6 – 10 psi added drop at 0.7 – 0.9 lpm 448G: DTC Cold Plate Simulation Results All cold plate scenarios evaluated for thermal and pressure drop performance PRESSURE DROP PERFORMANCE ACROSS FLOW RATES
p12
448G: DTC Cold Plate Measured Data Bench measurements for 500 W and 250 W scenarios validate the simulation model TEST set up
p13
448G: DTC Cold Plate Measured Data Bench measurements for 500 W and 250 W scenarios validate the simulation model MEASURED DATA — RESULTS TRACK SIMULATION EXPECTATIONS MAX CHIP TEMPERATURE — TEST vs SIMULATION MAX PRESSURE DROP — TEST vs SIMULATION
p14
• Problem to solve: • For direct-to-chip, analyze the impact of pressure drop across multiple cold plates. • Reliability testing at system level for to meet lifetime warranty. • Standardizing the cold plates for pluggable modules and present updates in OCP Cold plate workstream. • External cold plate: • Continue exploring TIMs that require lesser insertion forces, improved durability against shear forces (100 mating cycles) and lower thermal resistance. • To connect, please email us at: • divya.bandaru@te.com • sridhar.mukthineni@te.com Call to Action
p15
Open Discussion