Resource · Design Guide · 12 sections

AI Server Liquid Cooling Design Guide for Australia

Engineering design guide for direct-to-chip AI server liquid cooling — GB200, GB300, H200, AMD EPYC SP5, Intel Birch Stream. Channel topology, ΔP budget, manifold integration, CDU compatibility, AS/NZS context, validation methodology. Authored by ToneCooling Australia customer engineering support team in collaboration with the Huizhou design and engineering team. Read the guide below or contact us for supporting information.

Request the design guide
ToneCooling Australia AI Server Liquid Cooling Design Guide PDF cover
Process & standards glossary (AU engineering reader): TLP = Transient Liquid Phase diffusion bonding · FSW = Friction Stir Welding · CAB = Controlled Atmosphere Brazing · PPAP = Production Part Approval Process · ISO 9001:2015 (quality management) · AS/NZS 60068-2-14 ≡ IEC 60068-2-14:2009 (thermal cycling endurance) · IEC 62619:2017 (industrial battery safety) · AS/NZS 5139:2019 (BESS installation) · AS/NZS 4777.2:2020 (grid-connected inverter).
Page reviewed by ToneCooling Australia Customer Engineering Support Team · Carlton, VIC · Last reviewed: 28 April 2026 · About the team →

Resource Snapshot · 4 KPI

Sections
12
channel · ΔP · CDU · QA
AI platforms covered
5
GB200·GB300·H200·SP5·Birch Stream
Manifold spec
OCP DC-MHS
+ dry-break QD
Thermal envelope
ASHRAE W32 / W40
TC 9.9 Datacom

Engineering Reference & Standards

What standards underpin the AI Server Cooling Design Guide for Australian engineers?

The AI Server Cooling Design Guide is ToneCooling Australia's 12-section engineering reference for direct-to-chip liquid cooling. Platform reference: NVIDIA GB200 NVL72; manifold port-pitch: Open Compute Project DC-MHS specification; quick-disconnect coupling: dry-break type, specified per programme; thermal envelope: ASHRAE TC 9.9 W32 / W40.

§1What is direct-to-chip liquid cooling?

Direct-to-chip liquid cooling mounts a copper or aluminium cold plate directly on the GPU or CPU die surface and removes heat using single-phase liquid coolant flow rather than air. The cold plate is the primary heat exchanger between the silicon and the rack-level CDU. Practical thermal envelope for AI server platforms running > 25 kW per rack.

§2Which AI server platforms are supported?

Five primary platforms: NVIDIA GB200 (2,700 W per Bianca module), GB300 (4 GPU + 2 CPU kit, tested datasheet), H200 (8 GPU + 4 NVSwitch, 6.344 kW total), AMD EPYC SP5 (LGA6096, 350–500 W), Intel Birch Stream (Granite Rapids/Sierra Forest, 350–500 W).

§3What is the per-platform TDP map?

PlatformPer-die / per-module TDPTotal per kit
NVIDIA GB200 Bianca1,200 W × 2 GPU + 300 W Grace2,700 W
NVIDIA H200 8-GPU assembly700 W × 8 GPU + 186 W × 4 NVSwitch6,344 W
AMD EPYC SP5350–500 W per socket700–1,000 W (dual)
Intel Birch Stream350–500 W per socket700–1,000 W (dual)

§4How is channel topology designed for AI server cold plates?

Copper micro-channel (0.4–0.8 mm) for premium AI server platforms; skive-and-bond fin for cost-optimised volume programmes. CFD-tuned channel layout per platform.

§5What is the ΔP budget and CDU pump curve?

Typical ΔP envelope about 15–40 kPa at the manufacturer-recommended flow rate. H200 assembly ≤ 40 kPa @ 10 LPM (datasheet); GB200 ~ 35 ± 5 kPa @ 2.5 L/min (tested, TC-BZP-GB200-01A). CFD models the cold plate's ΔP–flow curve against your CDU's pump curve at the design operating point.

§6How does manifold integration follow OCP DC-MHS?

OCP DC-MHS r1.0 / r1.1 manifold pitch alignment standard; custom manifold orientations on RFQ for OEM-specific server programmes.

§7What quick-disconnect couplings are specified?

Dry-break quick-disconnect couplings, selected per CDU compatibility and customer programme.

§8Which CDU vendors are compatible?

Major rack-level CDU platforms are supported. Cold plate ΔP envelope tuned during DFM to align with CDU pump curve.

§9Which ASHRAE TC 9.9 inlet temperature classes apply?

W32 (20–32 °C) for H200, SP5, Birch Stream. W40 (20–40 °C) for GB200, GB300 with higher heat flux.

§10What is the CFD validation methodology?

ANSYS Fluent or Star-CCM+ commercial CFD; polyhedral mesh ~ 8–15 M cells per cold plate; conjugate-heat-transfer simulation under your specific coolant chemistry, flow rate, and inlet temperature window. CFD output = thermal heat-map + ΔP curve + flow uniformity map.

§11What does production-line QA include?

100% helium leak ≤ 1×10⁻⁹ mbar·L/s; 100% ultrasonic C-scan inspection of bonded interface; per-cold-plate certificates accompany every shipment.

§12What is the AI server RFQ input checklist?

Application, TDP map, target ΔP & flow rate, coolant chemistry, inlet temperature window, CDU model, programme volume tier, lead time, qualification scope, NDA appetite. View the full RFQ input checklist.

Compliance note. ToneCooling supplies cold plates that are design-aware to AS/NZS, OCP DC-MHS, and ASHRAE TC 9.9. Site-specific installation compliance is the responsibility of the licensed Australian electrical installer.

Continue Exploring

Related ToneCooling Australia pages

SOLUTION

AI Server Liquid Cooling

Direct-to-chip cold plate kits for AU sovereign cloud, HPC refresh, edge AI.

View solution →
PRODUCT

NVIDIA GPU Cold Plate Kits

GB200 / GB300 / H200 — copper micro-channel kits with tested datasheets.

View product →
CAPABILITY

TLP Diffusion Bonding

Core manufacturing IP — in-house joining for copper cold plates in heavy thermal-cycling duty. Eliminates brittle braze-filler interlayer.

View capability →
RESOURCE

RFQ Input Checklist

10-field checklist for fast Carlton-coordinated engineering response.

View checklist →

Request supporting design information

Read the 12 sections on this page. The request button opens our contact form; it is not a direct PDF download.

Request the guide
By ToneCooling Engineering Team · Last updated: